Speaker 9: what's again，

Speaker 9: there just was not damage of difference between the human written and airir and poll requests.

Speaker 9: in fact，

Speaker 9: humans are actually more likely to make p zero errors than devithin cortex or claud.

Speaker 9: similarly，

Speaker 9: the case with p ones where both devin and cotex created fewer p ones and p pets were even closer.

Speaker 9: once again，

Speaker 9: there just wasn't double ge vidence that the agent written ps were meaningfully worse or more critically damaged than their human wrt encounterparts.

Speaker 9: the next thing i looked at with the review rounds，

Speaker 9: a very common pattern for using rapile.

Speaker 9: is that you have a review，

Speaker 9: a poll request，

Speaker 9: give it some comments，

Speaker 9: and then you let the agent to fix those comments，

Speaker 9: make another committer in the poll request and usually two rounds enough for all of the issues from the poor request to be addressed and for the poor request to be grato to emerged.

Speaker 9: it seemed reasonable that a smarter or more capable author for a polar quest would be able to get to emergerable pr with fewer review round that seemed into it of you true to me.

Speaker 9: and so i started tracking average number of review rounds were deven little over two for people，

Speaker 9: two point two vercocotex to a half once again，

Speaker 9: entirely within arrange the mean number beerations.

Speaker 9: it took foreign agent or a percent to get a polar quest st.

Speaker 9: emergitble state was fairly consistent throughout once，

Speaker 9: again，

Speaker 9: no evidence of agentic pr mean worse.

Speaker 9: so now that there wasn't any strong evidence that these were quantitative，

Speaker 9: the worse or better i was interested in seeing there were patterns of failure.

Speaker 9: there were different in agent writain ps huhuwritten PS where they badred in different ways.

Speaker 9: and so i started tracking some interesting data around that here.

Speaker 9: re's how i did it.

Speaker 9: so group hyle leaves comments on all of these protocosts.

Speaker 9: it tends to identify all kinds of logic performance and security earars.

Speaker 9: and so i decided to run a very simple tech search and said，

Speaker 9: okay，

Speaker 9: are types of errors in different agents，

Speaker 9: different for different size.

Speaker 9: now the data set here is quite large.

Speaker 9: it's in the millions of protoquests.

Speaker 9: and so this data is reasonable，

Speaker 9: high confidence.

Speaker 9: and of course，

Speaker 9: you can't use all data.

Speaker 9: so i only used about two months of data.

Speaker 9: since prior of that，

Speaker 9: the agents are are essentially obsolete and IW.

Speaker 9: some very interesting datpoints clard for instances wanted to have times more likely than people to produce a equal injection error.

Speaker 9: author bypasses are half as likely coming from deven than from any other agent and from people and another interesting data point.

Speaker 9: but and plus one greos were much more common and cursor than any other，

Speaker 9: each of farm more common.

Speaker 9: it was interesting to me that there were this much variation in the qualitative aspects of what these agents were getting wrong when there were so much consistency in how good or bad they were and been measured in many different ways.

Speaker 9: so it might turn out that autonomous agents are actually quite good.

Speaker 9: and this is very exciting to me because it an an ttent，

Speaker 9: i kind of become programmmer before i had taken off.

Speaker 9: and i was used to，

Speaker 9: like many of us，

Speaker 9: the craft of programming，

Speaker 9: and it seemed like a lot of the vibe coded agents we were seeing in training，

Speaker 9: train four and training.

Speaker 9: trainfive were a lot more promised and substance，

Speaker 9: especially and applied to professional commercial settings.

Speaker 9: 怎么了。

Speaker 9: but the dejetory seems quite clear.

Speaker 9: these agents are getting really，

Speaker 9: really good and the're，

Speaker 9: getting really good at coding，

Speaker 9: not only in general environments，

Speaker 9: graplds customer bases made up much more of larger companies than a days of individual quotors.

Speaker 9: of course，

Speaker 9: individual quotors don't really review their code of don't nneed too.

Speaker 9: and so this sample said said，

Speaker 9: is essentially created compcompanies with well customers.

Speaker 9: and it turns out that they actually can do lot lot true vibe coding in a commercial setting.

Speaker 9: so we started thinking about what that means for code review and code validation.

Speaker 9: if most of the code is going to be written entirely by agents，

Speaker 9: and it seems like the agents writing pretty good code，

Speaker 9: then what should code review look like？

Speaker 9: i think a couple of things very clear.

Speaker 9: The initiating data point that emerges from my statistical analysis is that there was quite a bit of a difference between the 99th percentile of courts and the 50th percentile courts in terms of how much courts are writing.

Speaker 9: In fact, the median court that uses Graphite is writing fewer than 50 commits a month.

Speaker 9: The 99th percentile is writing close to 1000 commits a month at that scale.

Speaker 9: It is completely infeasible, especially if this is going to be the standard for how people code, to do anything resembling manual code review.

Speaker 9: It also seems not intuitive that we would be maintaining large test suites, as I go to hell.

Speaker 9: We got really excited about this idea that we could build truly autonomous code validation.

Speaker 9: We start thinking about what that means.

Speaker 9: And of course, code validation in the prior era was made up of these three general facets.

Speaker 9: You'd have testing, you have review, and you have quality assurance.

Speaker 9: But to us, these are simply implementation details.

Speaker 9: The goal of code validation was actually just to answer three very simple questions.

Speaker 9: Does this change violate the user contract? Does this change increase the future propensity of a user contract to violation? And third, does it fulfill what the author intended?

Speaker 9: And so we've designed Graphite from the ground up to exactly answer those three questions. We do this by reviewing the code, of course, in a more traditional sense where we use agents to look at the code, look at the blast radius of the change, find out the stuff that could be affected or related.

Speaker 9: And then we run the code in a sandbox, click around with web browser agents to try to break things, simulate user interactions.

Speaker 9: And that seems to us to be a much more effective path towards truly autonomous code validation because it's hard for me to imagine that this idea that humans would be reviewing large amounts of agent code, and that would become the new job of software engineering, is quite dystopian.

Speaker 9: I don't think anyone wants to wake up every morning and review slop all day.

Speaker 9: I'd much rather spend the time coming up with the genius ideas and letting machines figure out how to turn that into commercial grade code.

Speaker 9: That's my talk. My name is Ducch and you can find me at Rapal today. You can also come by our booth, we have an arcade machine. Thank you so much.

Speaker 11: Hi, I'm the CEO of Nori Authentic. We deploy an AI employee that understands your company, your codebase, docs, Slack and other kinds of data.

Speaker 11: We spend a lot of time thinking about how coding agents really work. Most people think coding agents only write code, but if you ask me, that's just bad marketing. Forget the name for a second. Coding agents can do almost anything. There's just one trick: you have to be able to think like an agent to get it to do what you want it to do today.

Speaker 11: Today we're going to talk about how we use coding agents to do something most people think agents are terrible at: make visual artifacts like slides, docs, and even video.

Speaker 11: Every day, the world pours something like 34,000 human years into making slide decks. Most of that time isn't the thinking, it's the fiddling. A deck that takes ten hours should really take about 25 minutes once you remove all the formatting and the branding and the moving things around.

Speaker 11: Say you need to make a slide. What do you do? You open a tool: PowerPoint, Figma, Canva, and then you start manipulating a canvas. Every one of these tools is built for human hands and human eyes. Click, drag, drop, resize, snap to grid, all motions and patterns that make sense for a geospatial view of the world. There is a data structure underneath, but it's in a format that only the application can read. What happens when you hand these tools to an agent? Well, the output comes out all wrong. Things overlap in weird ways. You can't see the text. There's no alignment. It's garbage.

Speaker 11: AI skeptics say that it's not just the tools, agents fundamentally can't reason about space, and there are whole benchmarks like ARC-AGI that are built exactly around that premise. There's a famous little test for this from developer Simon Willison. He asks every new model the same thing: can you draw a pelican riding a bicycle? But there's a trick. The agent is only allowed to use SVG. It's a quick gut check for whether a model can reason about space at all. Here's some examples of what the models actually give you on this test. And yeah, these are pretty bad, like genuinely deeply, really bad.

Speaker 11: So does that mean it's hopeless? Agents are just doomed to be bad at graphics? No, I don't think so. If you ask me, it's not the model. It's the medium. If I asked you, someone who is presumably human, to handwrite an SVG of a pelican riding a bicycle, you won't be able to do that either. SVGs are just a wall of numbers. You can't go from a wall of numbers to a pelican. You just can't see that way. That's just not how people think. We think graphically. So we build tools that let us draw on a canvas: Figma MCP, PowerPoint, screenshot replace ops. What do all these agent tools have in common? They all approach the problem like a human. But an AI is not a human. Asking an AI to use a canvas is like asking a human to write SVG by hand. It doesn't really make sense. You need to give the AI tools based on how it thinks, not in pixels. In language, words, token structure, that is its native medium.

Speaker 11: Imagine a language that's incredible at describing layout, that models have seen and trained on billions of examples of that. They understand intuitively that it renders to pixels and can run everywhere. All right, HTML lets a model think in structure. HTML tags have meanings built into the language: a heading, a chart, a grid. And the browser turns it all into pixels. So the model never actually places a coordinate, and you can get all sorts of visual effects, charts and layouts, fonts and motion, all of it for free.

Speaker 11: Remember that pelican from earlier? Now ask it to do the same exact task, but in HTML. Same bird, but now it's in a structure that the model can reason about. And you can read and theme and edit every single line of it.

Speaker 11: I've spent my whole life building slide decks with PowerPoint. So I always thought that those two things, slide decks and PowerPoint, were synonyms. But that's just not really true, is it? PowerPoint is a tool that you use to make slide decks. The deck itself, that's just the presentation mode. And as it turns out, no one in your audience is going to care how you got to the presentation mode. The editing format is totally arbitrary, so you can just pick the editing format that the agents are already good at: HTML. And if you need to render to a different format like PDF later on, we use this HTML trick to build all of our slide decks, our board decks and our sales decks. These are real things that we actually present and send out constantly.

Speaker 11: We use it for our docs, too. It gives our docs color and vibrancy all while following our brand. And of course, we also use it to make videos like this one. What you're watching is just HTML and CSS. It's literally just devs all the way down. Almost everything is better with a little structure and a little bit of color. Plain text is a choice, generally a choice of convenience, but it's usually the wrong one if you're actually trying to create something of use.

Speaker 11: Now I do want to take a quick break here and point out that a beautiful deck on its own is generally not worth anything. You still have to go and get all of that content, all of the things that actually populate that deck, right? So again, we can think like the model. If you just give the model access to your data, say your call transcripts or your emails, you can have the model build the deck from end to end. Let your agents do all the grunt work while you focus on vision and story. That's what Nori Sessions lets you do. I've built entire board decks for my company on the subway during my commute. That's our Nori, it lives in the fabric of our company. Of course, Nori ships with everything you need to make this all work. So don't bother reinventing the wheel. That's my little pitch. Thanks for listening.

Speaker 11: If you have just one takeaway, it's this: stop thinking like a user, think like the model. Give it the right language, and for graphics, all you need is HTML.

Speaker 12: Hi, everyone. Can you feel it? Hi. My name is Reren Nama, a mobile software engineer for the last fourteen years, and I'm here to talk to you today about tech stack reimagining, the mobile workflow.

Speaker 12: Say, you know, back in the old times, when a cursor was that thing you move with your mouse and AI agents were that dystopian character from Philip K. Dick books or movies, whatever fits your time, you know, just a few months back, back then, when we thought that we will still be using our IDE, just maybe slightly better.

Speaker 12: and now we know that we allow this AI to do like check-style engineering when we discuss we'd cloud called Codex, Cursor, whatever. And we just tell them what to do, and we don't use our ideas unless it's for the debugging or something that the agent couldn't figure out. And that in theory, should have made us ten times more productive, right? That's what everybody says, right? Are we ten times more productive? Do you feel it? I know. I can't feel that we are ten times more productive — not as a single engineer, and not as a whole group, and not as the whole company. So why is that? Why don't we see the promise of ten times more productivity come to actual life?

Speaker 12: So you know, they tell the story about how when factories switched from steam engines to electric engines, at first they didn't see that big of a gain. So yeah, the electric engines were better, they were more efficient, but they didn't see that ten times, twenty times, thirty times more effectiveness that they had been promised. And the reason for that was that they only changed the steam engine with the electric engine, but the real gain came some years afterwards when they understood that it's not only about changing the engine, it's about changing the whole workflow. Could you see they used to have like one giant big steam engine in the factory? And all of the machines were rearranged based on the power consumption and the proximity to that steam engine. So it wasn't organized by the workflow that they should have been like from the start to the end of the workflow. No, it was designed by proximity to that central engine. When they realized that, and they also realized that they could take the electric engine, make it smaller and put it inside each machine, and then they arranged the factory to make it work as the workflow should — because now it was made possible. Then the real gain — twenty times, thirty times more productive than they were before — not because of only changing the engine, but of changing the whole workflow.

Speaker 0: And that is what I want to talk to you about today. Let's think how we make things that were very possible before possible now, and we can change the workflow and then become ten times, twenty times more productive to do that. Let's look at the current workflows. The PM has an idea. They iterate with the designers, they iterate with the UX, they iterate with the dev, they iterate then back with the designer, then they iterate with the QA, and they iterate back with the dev. And maybe after all those iterations, maybe you have something in production. So what was that word that was repeating? So many times — yeah, iteration. And this is a problem because iteration creates friction. Each iteration creates context, which creates communication that needed to be done, synchronization that needed to be done. And AI didn't eliminate all of that. AI sped up code, but didn't eliminate the friction. It didn't eliminate the iteration. Why is that?

Speaker 0: So let us reimagine what we can do. Bear with me for a moment. What if, what if, what if instead of using one tool for designing, another one for testing, another one for coding, and then another one for releasing — what if we could use one tool, one code base? What if, instead of designing on Figma, then sending a design doc to developer in order for them to figure out how to bring those designs alive — what if designers could actually design on code and then send the developer a PR? What if QA could iterate with the agent itself, just getting a link with their simulator, and they can tell the agent exactly what to test, what to be cautious of, and if they find something, exactly what to fix? What if we could make the dev workflow work on the code itself? What if God was one of us? No, sorry, I got carried away there.

Speaker 0: And you're probably asking, how can we do all of that? One way would be to tell everyone to just download their Xcode and their Android Studio and teach designers, and PM, and QA how to build and how to test on simulators, and bloat their laptops with the 200GB of storage and whatever they do to their memory. That's one way. But let me guess that most of them would reject that idea — and for good purposes. So we can make another way. Maybe we just put it in our CI, right? So we let the agent iterate with the CI, so they don't have to download Android Studio, Xcode, and everything. But you actually know that CI builds take between 20 to 40 minutes, and we can actually let our agent wait for 40 minutes just to understand that the iOS code that it pushed actually failed to build. So what else? What can we use? Introducing cloud sandboxes. Cloud sandboxes are actually a concept that has been around already for many years, just not for mobile development yet. Using cloud sandboxes, you can tell the agent: here's a CI, talk to the CI, spin up a small VM that runs only for this iteration. The VM boots up in 30 seconds or less, make the build, show them a simulator under a browser in the cloud — Codex, Cursor, whatever. And then they can iterate over it, tell it to change that button, go back and test something, and change the code. And they push and open up a PR, and then the designer can work on code, send a PR to the developer. After they're done, developers make some iterations — run one, two, three different VMs to run in parallel. They send the PR for review. QA can take it from there and tell the agent exactly what to test and what to fix. And from there, it goes straight to the stores for review.

Speaker 0: So let's see how it should work. So imagine you see this screen. Imagine you're inside Cursor, for example. You have a chat interface to your left, you have the actual app to your right. The designer is iterating with the agent, telling it exactly what they want them to do, what they want to change, and see the changes immediately on the screen. Build time is faster — it's on the cloud and preview. Build time is faster. Then the iteration is not with the developer, but with the agent under the laptop without the need to install Xcode or Android Studio. And once they're done, they can tell the agent to take that code, open up a PR, and send it to the developer. This workflow is what makes us ten times more productive — not only because of using AI, but because of using AI to change the workflow. Imagine it and remove all that friction that we took for granted in the old times — that is how we become ten times more productive.

Speaker 1: Thank you.

Speaker 2: Okay. I want to tell a story about a factory that taught itself how to remember. Hi, I'm Rushup. I ran MachineCraft — a one hundred person factory in Northern Ireland. Steam? No, I'm a budget. None of that. And somehow, we ended up building 36 AI agents that run an end-to-end go-to-market. I think that's still a little ridiculous. Let me show you how it happened and why you can do the same thing.

Speaker 2: So here's the thing about our company from the outside, it looks like machines and metal. But the actual company — the part that matters — is in the machines: the knowledge of who the customer is, what we quoted them in 2019, why that one machine needed that weird custom piece. And for three generations, all of that lived in exactly three brains. Initially, my grandfather's, then my father's, and now mine — which is a genuinely terrifying way to run a company. When you sit with it, a lot of people have joined us. People have left us. That revolving door never stopped. And every single time, someone walked out, a chunk of our brain walked out with them. We weren't scared of the competitors. We were scared of forgetting — or waking up one day and realizing the whole company...

Speaker 1: Ah.

Speaker 4: It's great to be here. And today, with me here, is Gergely Orosz, author of The Pragmatic Engineer. And I'm excited to have a chat with Simon Arksson, a founder CEO of Turbo Puffer, a very technical CEO, and we're going to have a pretty technical discussion. But before we jump into it, Simon, I wanted to ask: where did you follow up with computers through PowerPoint?

Speaker 3: PowerPoint? I don't know if any of you know this, but in PowerPoint — you probably know this — in PowerPoint, right, you can make diagrams and stuff, and when you click them they go to another slide, and that becomes Turing complete real quick, right? You can sort of, you know, create very complicated, convoluted games. And at some point, you know, you make your way through the Microsoft Office suite, and you discover FrontPage — you remember FrontPage?

Speaker 4: I remember FrontPage. It was supposed to eliminate the need for all front-end developers — exactly.

Speaker 3: And it only worked in Internet Explorer.

Speaker 3: I remember a heartbreak. I had one day when someone opened a website I created in Firefox, and it just is all over the place. And then one day, I accidentally clicked the HTML thing in FrontPage, and it just showed all of this stuff that I couldn't make sense of. And it just started looking at it and then going online and finding little snippets that you could add in to make the cursor change in all of these different different different things. Then then just sort of escalate from there there, then you upgrade to Dreamweaver. And now you're coding. And then you're like, well, how do you make the pages dynamically, you learn PHP? And then for me, I exhausted the internet on Danish language programming advice. And I was I was around eleven or twelve. And so I just know know when got addicted to World of Warcraft for four years, but that gets you really really good at English.

Speaker 4: So you kind of start hacking get into deeper and a logical step would have been to just, you know, go to university and learn properly about this, but that's not what you did did you?

Speaker 3: I mean, I just I started just I mean, you then I learned video games that I learned English. And then you know, this, like massive arsenal of the web, I now would be very interesting because the algorithms who just speak Danish to me, and you could just you wouldn't have hit the wall like idea. So that would have been very interesting. Maybe I would been better at programming that would have been nice. And then yeah, then I just started picking up jobs and things like that throughout high high school when I was in in high school. Well, well, I got exposed to thing called the International Olympiad in Informatics. There is this thing. Yeah ah, and I had I had an internet friend, and she lived in Australia, and she was on the Australian team, and she told out there's probably something for the Danish team as well, but I'd never I'd never heard about it before. And so I founded on something like little mysterious website and then applied and then solve these programming problems that look very different from the html and php things that I solved until it was. It's like the algorithmic ish programs exactly sort of like this is not actually the kind of problem you would see there. But I think it illustrates about the the kind of problem that you might get right is you can imagine something. I okay, here's like N trucks. Here's M packages. Do M packages have these dimensions give me which trucks which packages should be in right, and then do something optimal. That's an NP complete problem. You can't solve that, but you could compete, but everyone else in the competition of doing the best thing. So these kinds of problems, right? Yeah. And so I started doing that in high school was working um I was working as well for I started. um and then I just Shopify found me while I was still in high school,

Speaker 4: And the whole I Shopify found me was it through your open source contributions. Was it? Was it something else?

Speaker 3: It was because I had written an article where I had I drop my iPhone. And I had, you know, iPhone is a lot like there. There used used to be a time and ever you drop your iPhone. And you just knew it was over for the screen. It doesn't really happened as much much anymore. Like screens have gotten a lot better. But back then it was like get one drop, and it was dead, and it just it couldn't use it anymore. And so I went back to one of these old Nokia brick phone. And this was back in 2013, and doesn't really realize all the pernicious effects of smartphones at the time. And so I wrote this article about how oh my god, I'm am like calling people. And I have my sense of direction back back and I wrote an article about it. And this article it went on Hacker News briefly, and then um New York Times decided to feature it no way yeah. And so a lot of traffic was driven to it. And then and I had Shopify recruiter put it all together. And um and I had a call with them. And then I don't think they realized that I was still in high school. But um but I had a great call with them. They invited me on site to Ottawa Canada. um I had no idea what Ottawa Canada is. I think the email said something like what's an Ottawa and no idea. um And so I went there and it was like walked into the building, and it was just a just felt. And and so I I interviewed with them. And then said what I got to finish high school. And then um and then I move to Canada um and to I it up up. Yeah, in 2013.

Speaker 4: Yeah, I think that's that's a like legit excuse for like not even worrying about college and a university. But did did across your mind,

Speaker 3: It did I thought I was going. I thought I I was doing a gap year I thought I was like, okay. I'm going to go work at Shopify for a year. And then I'll probably go back and look do, but I was just 哎呀, I was very insecure the time of of the fact that I hadn't studied computer science. And my only exposure had been all the IOI competition was a pretty good crash course in a lot of computer science. And if nothing else it had really taught me that you can just sit down and read a paper and just figured it out if you spent enough time on it. So I did I did that repeatedly. And in my first year year at Shopify, I just every time I heard something that I didn't know what was I noted it down in piece paper, and then I went home. And then that evening, I would just read about it because I felt insecure that, like well, someone mention she is mentions like TCP, surely they know exactly what's in the three way handshake and how TLS is like layered on top. And they've looked at Wireshark in all of that. I don't think that's true, but that's what I thought no. So I going to did that for everything that I encountered. Well, I was really good crash course. And then very quickly, it became clear that well, I just want to continue doing this. I don't want to go go somewhere else and and come back to this, because I felt like I already found what I wanted to do. So.

Speaker 4: Sounds like is a pretty accommodation of like you just having this like very, very national insecurity, then you're young, you know, you don't have the education that everyone else. And and inside company, that's just just dotting pretty like cutting it stuff even at the time and even even else today, right? They there, they're leading. So you just keep self teaching yourself like just catching up and go and let do understand that you just went deep in every concept that you understand. You didn't just just try to understand surface level, but like go as deep as you can search on the internet by books, whatever that is,

Speaker 3: I think it was just it. I just wanted to keep learning how computers work. and i think that this is something that i now look for when we interview engineers is that you just you can't help yourself, but trying to peel back the layers. And for me, that ended up with the infrastructure layer because was, you know, always process to the metal at Shopify. And i would just the set next to to at lunch because i was working on the on the product side, but it just i couldn't help myself. i was so i just wanted help learn what it was when they're talking about a reverse proxy. i'm like, why is it reverse? i still can't answer that.

Speaker 4: I mean, okay,

Speaker 3: you know,

Speaker 4: well,

Speaker 3: what's in reverse because it's proxy, right? I don't I don't know it's like an inverted index like what's inverted. It's like it's a terrible name anyway.

Speaker 4: I mean, it's still better one when you get to NAT NAT tables to look up. So some of those things like some of that. But I ar, I hear there's some like weird names with this. But at Shopify, what were some of the kind of like hard enduring challenges that use enduring challenges outages like like learning that kind of defined you that we're really also fun at the time or interesting to learn, but it would have been hard to get it elsewhere.

Speaker 3: Yes. So i think it was, you know, in the 2010s, there's like a bunch of SaaS companies that that scaled really quickly. and i felt so fortunate to have a front row seat to that. and so i ended up on the infrastructure team. and this was back in 2013, 2014 and docker was coming out. and so we were containerizing everything. and we were just every single year. we had to, you know the growth rates, but it's times times times quint in compares into the growth rates of companies today. but it was a company that was growing at you 120 140 percent year over year um and so every year, we were just preparing for a black friday. it was going to be a lot worse than the last. and this is back in the day of we're buying physical hardware, right? we have to like place in order at a particular point in time and do some interpolation based on that. and and Shopify also had to scale. and when you're scaling most software, a lot of the application layer problems end up back at the database layer.

Speaker 3: and so i just naturally found myself at this layer between rails and the databases. i didn't at the time contribute many patches to the databases themselves, but most of my time was spend orchestrating. so we were doing sharding because, as my my dear boss amillow used to say, you can't cache writes. so there's a fundamental point where you you just you have to move beyond a single shard. i joined around the time, and they did the sharding, and they did it think they did. they cut over a week before black friday, which is mind blowing, and very, but it worked. and then the the subsequent years we worked on things like going into multiple data centers. we also had this big, mysterious red red server that was like, you know, one hundred and twenty eight gigabytes of ram, which was a lot at the time today. it's not that much. and no one really knew what was in it. and then it went down one day, and people like well, that's super terrifying. because people had just been treating as this kv store. and so we started splitting it out. we did all this stuff around making sure that if you if you, if you go visit a shopify store, and the thing that stores your sessions is down, the right behavior is not just for the entire, the right thing be be down, but that's not of a default failure out, right? you're not going to rescue all of that unless you're in a programming language that really forces that. so we did things like build this matrix out of, okay. well, this service when this component is down should act this way. and i find myself writing the test week for a bunch of that. and then i was like, okay, well, we can't just mock all of this. and so i came up with this idea at the time. like oh, what we're going to do is we're just going to shell out to gdb and then into the process and then close the file descriptor that the database uses to simulate through the entire layer that the database fails. that was a little crazy. we never never to do on on CI, but it did uncover a massive amount of issues in rails to be upstream and things like that of round, just like handling failures at the connection layer. so then i moved on to create this proxy called toxiproxy. but you and you heard of this before, no, no, yeah. toxiproxy is it's just like a layer seven proxy that sits in between you and what layer four, but in between you and the database, but then basically have just this this proxy, and then mysql whatever doesn't speak the protocol. but you can you put it before, i can take take database down, make it slow and over time, it also added later, layer seven things of like do a bunch of failures this way. you're not mocking the low level drivers, but you're testing the drivers and their failure heavily as well. so then this entire matrix could be implemented in CI.

Speaker 4: so the basically the proxy was just like a really thin layer, which like was passed through. so it built the functionality to like simulate problems of database, or things like data corruption, or whatever you wanted to do. so you just do it in there. and then you can test anything that built on top of it.

Speaker 3: but but yeah and everyone, i call this proxy toxiproxy it needs to be in a layer exactly you do like, like, like you know, toxiproxy got mysql down, and it passes it a lambda of what you wanted to do, like, get this page do a check, check out whatever with the sessions table down. and this just uncovered tens of issues right in the mysql driver in the rails, like just like none and the ecosystem had been testing for this, and it was very difficult to see this in production, right? because they're mysql down. you're just used to just getting it back up and not like, what could the application actually have done?

Speaker 4: it is interesting. of course, we're going to talk a bit more about database, obviously. but just thinking about how a lot of the problems or some of the most general problems in our systems are always to do with state. and i never connected until now that i mean, state is usually there's a database. if there's no database, if you have state in the services, if you use memcached, you still have problems, you have network problems you have. i don't know, corruption whatever, but it's usually like more isolated. but basically, like if we have state, which typically means databases, if we have databases, and if you can simulate these problems, so suddenly you can, you can like predict a lot of things. so the problem with state often time is it's really hard to simulate problems happening ahead of time, unless you have the tools. so sounds like you were pretty successful with that.

Speaker 3: yeah, i think to my knowledge it though, running in like the CI system of shopify today. i don't know anyone in the crowd from shopify, but i'm pretty sure that it still does it. and so we wrote all these test against that to implement all of these different failure conditions. and it just yeah, it worked out great.

Speaker 4: so you spent eight years until at at shopify. so let's started from like, just a gap year. it just went on a year or the year and another year at what points did you think about leaving and why? and what was your kind of decision framework? it sounds like you were on epic, but not even today. shopify is doing wonderful. it's probably doing even way better than you know that growth kind of kept on. so i'm sure there have been an argument to stay, and you could stay in their organization.

Speaker 3: yeah. so i spent eight years there from 2013 to 2021. and i think it just came a point where i wanted to see something different. i've been inside of shopify since i was eighteen years old, right? i'd been seen one other start up in high school as like if i want to learn more about computers, more and learn faster, it might be time to inject some novelty into my brain. and so i left in 2021, and i'd worked on so many different parts of the infrastructure like caching. me and justin was now my co-founder rebuilt the entire storefront for shopify, which powers almost a hundred percent of traffic. eighteen months after we embarked on it. we worked on running shopify in multiple data centers. we worked on so many database scaling projects like caching all of these different things, right? a lot of the scalability came from the merchants launching lots of products on shopify, which would force a lot of traffic. but that's that's eventually how i left. and so when i left, i didn't really know what i wanted to do. and so when one of the projects i had, while i was at shopify, was this napkin math project, have you seen this napkin math? no. so napkin math was essentially just a table that i maintained on github of how much bandwidth can you drive to dram. what is a roundtrip to s3 cost? and how long does it take? how much bandwidth can you drive to an nvme ssd? how much bandwidth can you drive to an ebs volume? just a collection of probably there's probably like fifty of these numbers and then a ruby script that generates them all what all these things cost, what is a gigabyte of memory cost two dollars? what is a gigabyte of s3 cost, two cents? what is a gigabyte? what is this cost, ten cents? right? what is it cost on spot? what is it cost on a three year reserved? like what just have a massive table and create flashcards for almost every single cell. so i know of these numbers, and this was a project started to work on at shopify, because i found myself in this role a lot where i was going and review a project, right? so some product team would be like, okay, we got to do. we got to build this thing. so we're got to build this infrastructure to support a feature. and a lot of the times they would say, okay, well, we've gone and benchmarked database a, but the benchmarks are not very good. so we're going to go with database b and i hate benchmarks so much because that's not a satisfying answer to me. it's like this does not build my intuition. database a that you saying takes ten seconds to do. this should take ten milliseconds if you do a napkin math, right if it's a search query, right? it's like okay, you're searching for three terms there. each term has this many documents that match it. that's this many megabytes. we intersect these many, this many lists. you have dram bandwidth on multiple cores of a hundred gigabytes per second. this should take ten milliseconds. you tell me, the benchmark takes ten seconds. one of us is wrong either. there's a gap in my understanding, which is very likely or you have benchmarked the wrong thing.

Speaker 3: and in some ways, some reasons right is like okay you've done a benchmark. you don't didn't realize that your benchmark is doing a distributed query across a hundred different nodes. and so of course, the p, nine in nine, it's going to be really, really high right, unless you've cut that off, or it made some difference set of tradeoffs. so just found myself in these discussions repeatedly, where people were making infrastructure decisions based on poor benchmarks. and so i needed some, i needed some animal to go in and just be like, like, okay, we just do do calculation here here then then because i was always doing these, like little demos, or like writing little prototype scripts to to demonstrate this. but it was just, i just the argument of here's how a cache block looks. this is how many pages we have to visit. this is what a render node necessity takes. it takes one million second, you have to visit a thousand block, a block and then percent back. and the this is the difference to your query well like this is the query plan. correct? like, is there a bug in MySQL? do we have that desks like what's the discrepancy here? and i just got called with that bug. and so after i left Shopify was just writing a lot of articles about this. i was just like, well, how long this should query take. and then one hypothesis i had at some point, like, okay, well, how many writes per second can my SQL do? well, shouldn't the amount of writes per second? then MySQL do equally amount of things that you can do per second. that sort of make sense, right? every time you do a write. you fsync to persist a disk. so how many things can you do per second? well, one thing takes one millisecond should, you do a thousand write per second. well, that doesn't really match up like think a database can do more than one thousand, right? per second. why can you do that? so that was one of those things where i tested. and like okay, well, my YK​​O on on a little ThinkPad, could do ten thousand write per second. well, how was that possible? and now you're just ask, how was it possible because you batch? so when fsync happens on, usually a 4KB, yeah, right. but it's like that's not intuitive like it's actually i got caught was like, just like, you know, probably some, like twenty twenty to twenty four period, where i got obsessed with this question, where's like you're writing like the BPF traces and all of that to do all of this? this is like took me forever. and you and then i found out that oh, every fsync was like much larger than i would have inferred. like oh, it's batching you go into the code and you read it. and then you found some obscure article by it's always somewhere in like a central German town. that's like written some article about like how some intricacy of B-tree works and a patch that they did to. it's like the entire internet runs on small towns in Bavaria. i'm convinced yeah.

Speaker 4: and then you decided to start Turbular Buffer.

Speaker 3: yeah,

Speaker 4: how did you decide? did you know what you want to build there? was it more like i wanna build something something? database because you were clearly into databases. you done a an awesome job benchmarking, like, what is the theoretical like limits? you are very familiar. what this probably became like this expert in this niche and how i did you want to go to databases again.

Speaker 3: i think it was. there's three things that sort of came to a head. the project that worked on on Shopify was search, and i didn't have a good time.

Speaker 4: what did you use back there?

Speaker 3: i don't we don't need to names ok of other database companies. but was it was one of the like traditional search companies that a lot of different companies run? and it was very difficult to get it to do what i did. and i was just like the projects that touched that database just i couldn't get them to perform at the level, and there's there's no query planner. and like i couldn't figure out why it wasn't there and sometimes it track and then sometimes it really didn't track at all. and so i tried to learn as much as i could to figure out. and i started editing the source code of it. and i was just, i couldn't get it to track very often. it was very difficult to maintain. and so i just that was sort of like in the back of my head, i never thought i would touch that again. then the second ingredient in was the Napkin project. because it sort of just gave me a lot of facility with all of these napkin math numbers of what might be achievable with the machine, if you utilized it perfectly. and the third one was that doing this, you know, leaving Shopify in twenty twenty one have explained eight years there. and doing that time, i did this, i call it angel engineering. so i joined my friend's companies. and i just vested equity instead of just investing or something like that. and because i wanted to have my fingers in, and i wanted to see see what else was out, and that's why i left. and this problem kept coming up again again and it again and again, right? like ChatGPT came out in, twenty twenty two. and i was working with with a company then, and they wanted to connect a bunch documents to AI. and that's when the context windows were very really small. so you had to read every search very quickly.

Speaker 4: it was like a few kilobytes.

Speaker 3: it was a kilobytes or 4KB. depending on a model was very, very small. so you had to reach for search, very quickly right now. and, so i work with them. and i was i was i create a little recommendation engine. and the recommendation engine was actually quite good, like i found out that one of the cofounders wife was pregnant through the recommendations that i was getting when was running his feed, but but was was was wed, but he was recommended. yeah, i mean, it was just like ah i was reading like ah i did get permission. i just like i don't get anyone expected to be good enough. and just like, okay, it's this. this thing is just king, and then i ran the back of the envelope math on what it would cost to do this for everyone. like all the users. this is a company called Readwise is like articles that you save and then and insert later. and it was going to cost six grand a month. and this was a company. it's a bootstrap canadian company. they spent about five at the time they were spending five, five a month on all the other infrastructure combined. so just it didn't fundamentally in a company. if you're doing an investment, you have to to run some growth margin on top of whatever you're paying right, it just didn't line up um and so we just didn't ship it. and i worked on, i'd know tuned to autovacuum on PostgreSQL or something like that, which is a good past time. and then i just just couldn't stop thinking about why i was so expensive to store all of these vectors that we were using for the recommendations. and i suddenly did the napkin math day day. i can we just use it out of S3 and do some clustering and then organized the files and just the way. and like maybe you could build that. and then one day, i just gonna said, fuck it and did it. and like sat down and started to like to write it out. and i spent the summer of of twenty three, just hammering my head against the wall, trying to find an approach. that i can get the latency that i wanted um because the bandwidth with S3,

Speaker 4: is it as a really good bandwidth but latency were talking hundreds of milliseconds or right?

Speaker 3: yes, the p ninety nine on it because two fifty six or five twelve kiloby object on S3, um a two hundred milliseconds and and your same p, nine nine.

Speaker 4: because like when you're talking large scale, you wanna care about about p ninety nine,

Speaker 3: right?

Speaker 4: think think think when you're not talking my p fifty,

Speaker 3: when you're designing a system, you want to optimize for the p ninety nine. and especially because when you're deciding a system on on S3, generally in every mill, second, you're not doing one request, you're often doing lots of request, right?

Speaker 4: you're going to have the p, nine,

Speaker 3: nine real ick exactly. so it's like if you're navigating a tree on S3, right, it's like kay, you get the upper layer of the tree. two hundred mill. second, you get like another layer of the tree, two hundred mill. second, you get a bunch of leaf of the tree and two hundred, a mill second. so an aggregate, you have like you want to look at the p, ninety nine, probably even the p, ninety nine point nine to design the system properly because you need to minimize the number of roundtrips that you had to make.

Speaker 3: so i just didn't sketch that out and try to bunch a different approaches.

Speaker 3: and then and then finally, in july of twenty three, i got something and it seemed to work and then rewrote it probably twice.

Speaker 3: and then released in october based on just that summer and then working through it and you built on on top of it.

Speaker 4: because guess your ability and all of and just really good, how do you make it fast?

Speaker 3: we didn't in the beginning or i didn't in the beginning. it was just me at that time, and it was really like it was a project. it was not a company. it was to satisfy my curiosity. it was like did not set out to do this, like i'm going to go like raise ten million dollars. and i was like, i barely knew what AVC was like. guy was like, i just had to do this thing. and i was so focused on doing it. i'm so clear to me that if i wasn't going to do it, someone else was going to do it. and i just became fully obsessed that summer with it. and so the first version was the simplest possible thing. i think i am a very pragmatic person like i didn't get buried. i barely read like literature. i sort like, you know, read enough. it is that got to the basic idea barely implemented because that would have taken too much time. it was the simplest possible version of what it could be like, really.

Speaker 3: what you have to imagine is that the simplest way you could do this is you run some clustering algorithm on the vectors, like you put the clusters, and then you put the clusters in files. the files are called cluster one, cluster two, cluster three up. and then you have another file called centroids of the clusters. and then you do the search by downloading centroids, looking at the centroids, and then downloading the closest clusters. there's a few optimizations around merging some clusters that were in files and so on, just to control some costs and some performance. but that basically and then getting that to scale. that was the first version.

Speaker 3: well, i didn't need to implement a caching layer. i just put the reverse proxy in front of s three with nginx, and then had a caching on that reverse proxy. i like, i do know what it is. i just still don't know what the reverse is about. but anyway, the reverse proxy, reverse things. and the performance in this case, um baby. that's what it's about by caching, right? all of the s three objects. again, it was simply that. because like i'm just going to put that in front, i knew how to configure nginx. like i've written more nginx lua than, yeah, a lot of nginx do a very good software. um just had that cache in front. and then the way that i would do things like deleting in the cache was just like shell out to xargs and just removed like things in the cache of reverse proxy, the directory structure on nginx. and that's what we shipped. it was just running on a single server and t2 micro instance. i was like, okay, let's see if anyone gives a shit.

Speaker 4: yes. so so far. i mean, this is kind of like cool engineering and like a cool side project and like a bunch of novel ideas and i don't know, like i think just some solid engineering. how the curse are come into play. because like when i learned about turbo offer, i was saw them with cursor about like how they built there. they're back in their database, how they scaled and tell me all these migrations. and they're telling me like, oh, so we were on postgres, but it didn't know they did something else and postgres it. and we worked out well. they went eight months of turbo offer and which is a managed service, a postgres, and it didn't well and which is very surprising. and they like oh yeah, then we went to this thing, all turbo offer, and they worked well. and i was like what turbo offer, i think oh at turbo offer. i well think that we were around their first customers. and this never computed to me. cursor was already massive about that point. how did you meet the folks? and how they become were the first customer.

Speaker 3: they were the first customer, the first, the first. no, um build this out after i just launched on on twitter as okay. i build this thing. and frankly, it was like, hey, you know what thing to me is like i am so sick of working on this. i guess, like i've been working on this all summer. i don't know if anyone cares, i only want to work on this. if anyone cares, let's put it on twitter again. single t2 micro instance on a corner somewhere in gcp is like if someone goes surprise, i'll set it up properly on multiple. and like i'll just block on that. but let's just block on that. let's like the MVP of MVP. anyone who's actually worked in the internal on databases would never have had, like would have had too much pride to ship anything like that. um and i'm just you know, i've worked on. i was just releasing it like a saas project. why can't you work on a database like it's saas? i don't know. it's like if anyone uses it, we'll do a proper. i know how to run software with a lot of nines. and but was not not a problem. i was like, it was very, very. it was the simplest version of what it could be. and then i released on twitter. i was that you could do a million vectors for a dollar. and before that, i think the cheapest was maybe hundred dollars per million for something that actually work. yeah, um and i knew it was reliable, right? i knew like i had had these invariants. like if you shut down all the vms, like no data is loss, like all the writes are committed directly to s3. like it has all the same invariant it had today um and cursor reached out and knowing them. now i'm sure at the time the cursor was maybe people and and knowing the founders. now i am sure that they had set at the dinner table one day and were like the unit economics are what we have right now, where all the vectors are in dram are not working. why hasn't one built it? where we can put in memory the actual codebases that are actively being used. we can put in memory and everything else just sit in object storage. and we just hot loaded in, it makes so much sense, right? you open the code base a few seconds, and it's in ram and then the queries are as fast as anything else, it made so much sense.

Speaker 3: so i mean, at the time they were, if you look at some of the one of the co founders early tweet, he talks about using history for KV caching. and they did like that, which barely anyone is still doing even though yeah of price. yeah. and it's very, it's very uncommon, and i think it will happen right, but they were ahead of their time, even they i think i don't know there are thinking of building it themselves. i think that's quite likely um and they found turbofer, and it just perfectly pattern matching to that again. i don't know if this dinner conversation happened. or if this was just inside like pattern matched. and so we exchanged a bunch of emails, and then something compelled. i didn't know anything about b to b sales. now i love b to b sales. now i don't know anything. i was just like i just want to help them because they were. they had some unit economics that didn't line up. so i just went to sempazis go. i went in canada. i went to sempazis go, and i showed up the office. and when i showed up the office, they um they were having some postgres problem that they were discussing.

Speaker 4: yeah, the end of several problems.

Speaker 3: yes, yeah, early on. and i was like, oh, you guys have PG analyze, and they said, oh, no, we don't and let's let's let's get that going right? let's look at it. and it was the same thing as it always is with postgres, which the autovacuum hadn't run enough. and so they had all of these, like index bloat. yeah, they should be doing index scans and level about. so we were talking about all of that and just helping them. it was like my database geekiness just kicked in. and i think this build enough trust with them that okay. well, maybe if he knows how to help us with the database, maybe he also would know how to build one. and at this time, i also approached i thought was the best engineer who ever worked at shopify, my cofounder justine and she came on. and the first thing that she did was um remove the reverse proxy nginx cache with a falbase cache, just the direct cache layer.

Speaker 3: great like the yesterday work. and so she was online, she was starting work on it. and um and cursor cursor cursor, then that night was like, okay, well, we're going to migrate. so so migrated everything over the course of a week or two after that, um and cursor was a small company back then, right?

Speaker 4: yeah. and they were just in beginning of their massive rapid growth exactly.

Speaker 3: and I told them that i was going to reduce the bill by ninety five percent. and i did like we did just see, and i did we like they came on and their last bill with their previous vendor. and the first bill with us, it was ninety percent lower yeah.

Speaker 4: you're nice for not naming vendors, but but i can see it as they talks to them. and it's in the deep dialog cursor. it was it was eighty or was or war specifically.

Speaker 3: so this was this was not this was not postgres now.

Speaker 4: but yeah, but there were the reason they went. there is reliability was their main selling point i'm sure the unit of comments, what would have been there, but he had this was, and then what Swallet told me is, he said, like look like there's a few things that we did. it should never ever do. and he said of one of them, you should never ever bet your business on a tiny startup where you are. there are only our biggest customer except... and is that i love the guy. so i guess it just comes to show that even in your... to me, what this story is shows us is you can do things. and when you build high quality things and you're pushing for things, good things can just come to the customer of cursor. when you're startup, it's okay to take sometimes your rational risks when you have conviction. and it sounds to me that you gave them conviction by showing up in person, by helping them by showing that, you know, yourself, like you, suddenly brother, your your tennis, or eight years of shopify experience in your curiosity. and they probably took risk because of that, not because you were some your dom vendor, they probably never never done that.

Speaker 4: so the fast forward today, turbopuffer is now a lot bigger. you're working on some cool things, but you have this very interesting business where for you cpu are important, right? you run on mostly cpus, and you told me a story over dinner yesterday that you met jensen and jensen. he really wanted to sell you on gpus. can you tell me how that meeting went?

Speaker 3: um yeah, jensen huang, right? yeah, i just i never met. i never never met jensen before, we were at an event and at nvidia. and we're just doing presentations in a big HQ super impressive. yeah, exactly they invited a couple companies to go and talk about our businesses and how we can partner with nvidia. and i don't know, i was like, i think i was in a goofy mood that day. and so i went up on stage. and i said, hey, i'm sven from turbopuffer. yeah, if you're wondering about the name, it's if everything goes south, we can always pivot it into vapes. i was kind of nervous as this is what i said.

Speaker 4: and then and then he said back to what who was in a room as a jensen was a director or it,

Speaker 3: then it was jensen. and then he has like, i don't know, it's it's fifty direct reports or it was like, i know there was no answer, then a bunch of the like nvidia leadership, right? because you go there. and then you talk about that, you find about... the partners are more together, right? and so i said, yeah, this is also plenty of equity that we could have an into vapes. and then he said, i was already nervous. he said, judging by your slide, maybe you should.

Speaker 1: he did not,

Speaker 3: and i didn't know what to say back to that. so, i said, well, jensen, do you...? he didn't answer the question, and someone on the team wrote to the whole company. terrible poffer simm just asked jensen a few vapes. um and then you know, this is a great start, right? and then the team had sort of talked to me beforehand. like sometimes, we got to make sure we don't say the c word.

Speaker 1: we can't say cpu.

Speaker 3: so i just couldn't stop talking about how cpu was like amazing. it's so sick like we love CMD and up like like there's so many cpu. there's so easy to get like it's just the right in cpu land like you don't like, i think i stop short of saying. i'm so glad i don't need gpus, but but it was just i just couldn't stop talking about cpu, yeah. and so jensen took an interest in that.

Speaker 4: so okay, who knows? like i'm i'm sure you may you made an impression. so maybe he made it his mission to, like at some point, get you guys onto GPUs. but what is thinking of cpus, can you tell me? but what you're seeing inside of the hyperscalers, the cloud providers you're now in aws, you're in gcp, you're on azure, what i would think now. heavily is there's a GPU shortage. and when i talk with inference companies, they are and and AI apps, they're getting whatever they can get. no would think getting cpus is should be easy. is it?

Speaker 3: no, it's not anymore. what was happening? okay, tell us about the dynamics on the ground and what you've learned?

Speaker 3: yeah. so i think that gpus will probably continue be scarce like, i don't know, maybe there's going to be some surplus. i refuse to speculate too much about the macro, but i think as as RL is becoming a very, very large amount of the workloads that needs a lot of CPU. so the labs are sucking up a lot of CPU because you need CPU to be like, okay, we need to like teach this model, how to how to search. we need to teach it how to use graph. we need to teach it how to boot up bash. it needs to run all these things, things learn from that. that takes a lot of CPU. and so i think RL is consuming a lot of CPU. and then also just all of the agents are running on CPU, right? they need to do all the very general purpose things on CPU. and so as the demand curve is shifting to the right and that becoming more and more applied, and that feeds back into RL, by the way, right, because of things become more applied. like oh, the models are not that good at cat or ship building. i don't know, and then you know, you have to spin up even more, RL environments to do that. so think that's what we're seeing. and so we're on the other end of that needing these CPU. we and a lot of NVMe as well. um and a lot of this right now is tied up in dram, right of where, like you need a lot of also for the GPU servers. um but i would assume that it gets a lot worse before it gets a lot better on the on the CPU side. um and i think even the big companies are fighting among each other right to get cpu and even know we're selling to companies that we also fight for cpu with and against right even. it's it's really difficult. and so you write things to try to make sure to get these cpu as fast possible.

Speaker 4: yeah. and yesterday, i was at dinner that you was with your team, where where actually a bunch of our customers, a bunch of them, our AI labs or AI startups. but a lot of them, one of them, uh, reflection had huge massive footprint. and they were telling me that they're in a situation where they cannot buy more like they bought all of the gpus or cpus. they max out. they have longest contracts that possible, and i didn't realize how competitive it is in the cloud when you go beyond a small fish, like a medium sized or even a large fish. now like it's it's interesting. so now you have this even you're having this kind of fight behind the scenes that is maybe not as visible exactly.

Speaker 3: and i mean, you, you work with the clouds, right? you work with them to talk about which regions have have cpu, which regions are getting. it comes down to power right of like, okay. well, where is the power, which is generally where they're going to ship the new cpu. and so we have to work with some of our biggest customers on that. these are real constraints right right that are making their way to us. we're just very fortunate that it's very easy for us to run lots of turbopuffer instances, because all we need are like like a few cpu and the memory, and then s3 and then we're in a good place. but there's lots of change that we can make even into the architecture, but try to protect from a lot of this.

Speaker 3: now i'd rather spend that during effort on other things, but we are very, very good at using a lot of very different skills, right? so we don't need everything to be a particular cpu or instance type. we can run with many, many different types of machine types on a few meaning.

Speaker 4: that's like it's an fancy name for like different different machine types.

Speaker 3: exactly, like you know, feaning that's or ia g or whatever they called, what's your favorite one? um we really like right now, the um C3 on on GCP. um does that for, these or also performing really well now that we've done a done a bunch of um of optimizations to them. um those are really great, great machine types we really like those. and then the arm c forch es as well um on on GCP of we like those. but i think that in general, like when yeah, a bunch, and we very easy shock of a bunch of, but that shop pify. i was also part of deciding i had ABFCM right a few months out. you have to tell the cloud providers, how much you're intending to use do commits on all of that, right? the clouds are not infinite as they seem, and you're small.

Speaker 4: and one way, of course, so like get like infrastructure. and and also just like credibility is venture capital. if you raise a hundred hundred million dollars, a billion dollars, some of your customers just raise two billion dollars. actually, i talk with them yesterday, you know, i gives you credibility. you gives you cash, you can pay for this thing. um, your specific turbal profers, your relationship to venture capital seems very interesting. i never heard you announced a ise until maybe just very recently. can you tell me how and you told me that when you start this thing, you didn't think too much outside of just buildings and ooall stuff. how did you think about venture capital? and how do you think about raising because i can i feel you, you have a very freshand different perspective and which which is typical inside of siloammi.

Speaker 3: yeah. so i think to understand my how i think about capital, you have to go back to the the beginning of turgo puffer, right, where i promied cursor that just seen. and i could get their bill to four quay a month. and this was based on some very rough nafkan math on. okay, if turturbpuffer was a better implementation than occurrently than then, it could cost this much. and that's the pricing we ship with, and that's will be guaranteed guaranteed curser. but the suoftwer was not that good, like it was very reliable, but it was very simple, right? that's like a core engineering principle of me. me. it's simplicity above everything. you and i have talk before me, me, how suffered age is well and some of the advantages of seeing um you silong tenteners inside a company, you you at a long tender er, uber at a long, long tener tratrarififiso. you see sisimplicity just almost always wins.

Speaker 1: and at the time, i was not convinced whether this was was venventure skill opportunity, because i understood that if you take venture capital, no matter how many smiles are in the room, everyone is sort of expecting that you have to earn a big return on that on some timthat that makes sense to everyone involved and everyone involved. are you pension funds in canada like that did like it is like a whole stack right of of people that they need to. so at the time that, like you know, i don't know if this could be a builility. i didn't know that in the very, very beginning um it wasn't completely clear to me. it felt like a very nich kind of product right to build this particular search engine. and that was completely fine with me. so you know which it was fine. and so then i just looked at the curser bill. and i look at my GCP bill, which just what we started on and you know was was like you dumb danish person, who is just like, okay, like this number should just be lower than the other number. yeah, that's sort of you know. and it's just i don't think i spend enough time and sanfron csco because i think the money over here, it works a little bit differently. that's just that's all.

Speaker 4: i knew you you're doing business one or one as as long as you're making a profit, you're good right?

Speaker 1: yes, that was was i'm i'm not kidding in this exaggeration that it was just like that just made sense to me that just even and i were just going to go optimize this until these rutt numbers were roughly equal. um maybe if if we could get some of other workloads, we could start paying ourselves, but that, that was like very much the philosophy at the time. um because i didn't know if i could go raise a bunship of of money. i didn't know anyone who had the money. i didn't have any relationships. um 're annxlual outsider of i was an outside der. i was that was like outside square, right? i grew up in all who's denmark, and i um i dn't move to ottawa canada. it's like i'm an outsider to canada a in canada and outoutsider two anfr's disco. so i was just thinking about this from first principle. oh, h, your eventure capital, you need this return you needed on this timeline. i don't hope i can deliver that yet. i would need more data to decide that because i want ted, i kind of want to keep working on this. and now i have to get to this point for it to not be a failure in in january. then i there's a person that i was at III with in twenty twelve and twenty thirty, and his name is byn, and he was on the northermasasdononly teteam at iri, and he was his. he was really good. he was so good that the north massaony an team called him god. i don't know why, but that was what he went by, and he was yewas very good, good, good rouroup. i really want want to work with borin, but i couldn't afford to work with boys, and he was very much like this is what i could live off. you know, i just got like i want to build the state like that would be like this is what it compute. but at this point, just seeing that, i hadn't take ken a salary for like six months and 'd already. we'd already spent like tens thousand hundred dollars on like on GC people and all of that. and just like i don't think we can on GY people and can do it. and so i had met one one individual in in silicon valley. his name is loccky, and and just i ended up just calling him and saying, hey, i kind of want to learn a little bit faster here. can i can we raise like seven hundred kay? that's like what i wanted to raace. so it's like i wanna have two engineers for the rest of year just seen. i still need to be paid. and then a little bit of buffer room. it's this is what what need. and and this doesesn't pm f, and it's a big opportunity by the the year. i don't think wegoing to bother ther and will just shut the whole thing down. and we i want to have a taking a time will return everything to you 啊. i think there was the first time you heard anyone said like that. um and i told some other vcs that at the time, and that was terrifying to them. i think to someone on the west coast, this sounds like you have low ambition or something like that. um and to me, it was just like i don't know, just came from a when i don't know how to play a game. i just play it open cars like this is how i see it. so and we were it was very clear to us that we wanted to do this. and but also, it became clear to us that we didn't want it just like keep working on this and leesit could become big. and you were starting to builp condition conviction that is actually become really, really big. and so we did that and hiirer born, then became profitable later that year. and then just continue to higher. and then it's like to raise more money. you need sort of there's six reasons to race capital. the first reason to raise capitals to fund r and d. um that was the reason that we raied capital in january because we funded RND with a lot of our own. you know, opportunity cost sts, not taking a salary and then paying the bills ourselves, but we wanted to learn a little bit faster. and so we hired boorn and morgin us the first engineers. and in the second reason to raise capital is the fund growth. you build something, and you want to tell the world about it, and you want to spend more capitals to do that. the third reason to to raise capital is for the founders ego. it's a very popuposly. it's very popular, very, very popular, right? big numbers, lots of press like. and i think this is a very, very dangerous reason to raise money. and i wish that it was more talk about because you're deluting all of your employees when you do it, you are um setting a certain price for future employees, and they're upside. it's it's for some people who can become a status game. and that's not what it's about. we're here to build a big business together, and this is not a reason to raise money.

Speaker 1: but i do think that it happens um the fourth reason to to raise capital was to reward your employees, right? it's if you're on a very long journey. and you want to work with the best people in the world. and by definition, there's not that many best people in the world. so you want to reward them. um that was the reason that we took more capital in december um was to allow the employees to liquidate some equity, instead of waiting for some like event like an ipo or something like further out. um fifth reason to raise for a strategic partnership. there are strategic partnership that have been made in this in this city that have made companies. the six reason to raise will be doing ma or or something like that. but it's like you have to be very honest about what reason you are raising in those six. first reason we raised was one. and second reason we raised was four.

Speaker 3: um which once the first reason to raise was r&d, and the second reason was to provide liquidity to the employees.

Speaker 1: yep.

Speaker 3: i think it's a nice and healthy way. and i think the equity part, we don't talk about in the identity, and so especially the the founders who are to to take you, because this seems where a lot of people are raising. it will be part of it as closing thing. i want to ask you about the way you have a remote culture these days. i'm seeing it especially for companies that do anything with ai may that be building ai infra or or or just ai products, a lot of them prefer in person having a hq often times in sf or wherever your headquarters is maybe be london or somewhere else. because you often, these companies often find that they have faster iteration is just a few layers can of between. and of course, speed, as is very, very important. you have started remote. you're still remote. how is it working? and what kind of quirks are like or tribal ways have you found to make this work better?

Speaker 1: yeah, i think the company started in twenty twenty three, sort of like on the on the post-covid era where a lot of companies were just remote. um the shopify infrastructure was remote since the very, very beginning because it's not very difficult to get them to move to ottawa. um and so it was natural to me is like, okay, i think there's kind of maybe two cities where you can build a database company fast. and that's san francisco and maybe new york there, maybe other cities, right? but that's like kind of where it's been done. and so if you don't want to do that, i think you have to go all in on on some distributed model. and so we've had to figure out what is that distributed model mean for turbopuffer. it doesn't mean the absence of in person. we get everyone together twice a year in some in some location. um this year, we were in south africa, then we in mexico city. and so so it's like that's that's not that often. but one of the things that we've been trying to do, we have this concept called camp fires. and the concept of the camp fire is that when a couple of people just sort of randomly congregate in a place, you call it a camp fire, and you encourage as people, people as want, come come and join. so for example, this week is a tribe of of camp fire in san francisco because i am here for this conference and a bunch of other things. and so everyone is invited to come like we're going to go for customers, right? we're going to put on dinners for our customers and things like that. and we just make a thing out of it and and spend time together, and we encourage everyone to come. we've also gone to the extent now of we won't encourage that, but not everyone, not everyone needs to go to to campfire all the time. some people just want to, you know, lock in and hack into the terminal. and that's great. we have people that just make it to the offsites twice a year and otherwise they're home with their families. they don't. they don't spend time on an airplane. um fantastic like that is completely compatible with this model, and there are other people with the company who are on a plane, probably every two weeks. um we had someone the other day where they saw, a camp fire happening in new york. and everyone was dialing in from a meeting room in new york, and she had so much fomo that she took an uber straight to the airport in ottawa, flew to new york to hang out with the team, right? and i think that's fantastic um and we've also introduced these things where um if you you do a conference talk or a blog post or something for turbopuffer something a bit extracurricular, um we give you a turbo credit and a turbo credit allows you to upgrade your next flight to business class, which, again, encourages spending time together with the team. um and now i mean, turbo credits are probably going to take on a life of their own. we're almost talking about doing a central bank and doing interest rates on the turbo credits and and doing a betting market on the turbo credits. and so like this might take on, it's life on its own. and you, you know, if you're at a conference like this, there's some of the our engineers here who are just to interact with customers and beyond, like standing on a like expo floor all day is quite taxing. and so if you do that for two days because you want to do it well, you get a turbo credit, right? and so it's just like these fun little things that we try to do to to encourage people to meet if they want to meet.

Speaker 3: thank you. well, in this session, what i found very interesting as turbopuffer is so many ai companies are using you as infra layer, but in this conversation, we managed to talk very little about ai, a lot more about enduring principles, pushing being curious and the human connection, how important this is for people to work together to trust each other. so just thank you very much for this. so let's give a big round of applause for simon.

Speaker 1: thank you so much.

Speaker 8: supertengines lab and infrastructure. today, we're going to be talking about production with agentic systems. when most people hear the word evaluation, they think about benchmarks a model scores ninety percent on benchmark, a new version scores ninety two percent, a team celebrates. but agentic systems have fundamentally changed what evaluation means today. the system reasons, plans, calls tools, retrieves information, executes workflows, and interacts with production infrastructure. the question is no longer did the model generally give the right answer. the question is is the system quality good enough? i would like to discuss how evaluation is evolving for modern agentic systems and evaluation infrastructure. this is the problem almost every ai organization is encountering today. offline benchmarks continue improving, yet production reliability often remains unpredictable. why is that? because benchmarks measure model capability, production measures system behavior. a benchmark doesn't capture tool failure, api context changes, data availability, long tail workflows. and as systems become more dynamic, the gap between benchmark performance and production performance grows. the result is what many teams face.

Speaker 4: all right? hello, everyone. my name is kevin. i'm going to be talking about antigravity. so are there any world cup fans out there? whoa! imagine you are coaching argentina and you're in the eighty ninth minute. and you have messi on your team. what play are you running? it's called give messi the ball and get the heck out of the way. llms aren't just role players anymore. they can be your star player, if you build the right product around them. and to let your star player cook, you have to get out of the model's way. we might want to get the slides up. oh, they are great. um so antigravity is google's agentic coding product for technical and non technical users. we launched back in november of twenty twenty five and have been accelerating adoption, both within google and externally ever since. my name is kevin, and i lead part of the engineering team on antigravity. so let's talk a little bit more about what antigravity is we have and always will be unapologetically agent first. so we debuted the antigravity ide last year with a brand new agent manager concept, and it was a platform to manage and orchestrate many agents. since then, we've actually extracted our agent and launched our own antigravity cli. and last month at google io, we had the pleasure of launching antigravity 2.0 and in the theme of getting the model out of the way we actually decoupled the ide from the agent manager. so now you have two separate applications, and now you can use the agent manager in a standalone app. and so a picture is worth a thousand words. here's a screen shot of antigravity 2.0 in action. as you can see,

Speaker 4: not only is that your own dedicated mission control for your agents and projects, you have sub agents, you have all the new models. you have work trees, scheduled tasks, voice mode. there are so many things to impact with the product, but i don't want to spend today telling you about the products. i want to tell you a little bit more about the behind the scenes, some of the principles that went into it. and notably some of the things that led to its road map.

Speaker 7: 嗯.

Speaker 4: so as some of you for the long time, AIN ge ans ans, this is actually my fifth time speaking at AI enge. um and i've been building developers tools since twenty two. and the one thing that has stood above all other lessons that i've talked talked about is the idea of scaling with intelligence. this means that as the model gets better. so should your product in the frontier edge of whatever model you are serving should be appparent inside of your users product experience. so let's get into more concrete examples of what this means.

Speaker 4: so for those of you that follow me on x or hear me just gap generally for the last four years, you'll know that i've been working on a number of these sort of transformations year over year over year. in twenty twenty two, i was working on auto complete and chat cybars. this was based on embeddings rules files AST syntax, three parsing, basically everything inside of that app is deterministic because that's all of the model could really handle. and in twenty twenty four, when agents came onto the scene, it completely chanchaned. how developers were going to do work with. it came new permitives like mcp custom tools and permission systems. and with twenty twenty five, we introduce antigravity is agent manager with many other products following suit in that similar forrein factor with users managing many agents at once in parallel. and this led two things like skills, hooks artifacts in a couple other primitives, and that sort of defined the twenty twenty five era.

Speaker 4: so let's talk a little bit about twenty twenty six and what those primitives might be. before we answer this question, i want to take you back to some of these battlescars that are a little bit closer to home. scaling with intelligence really is not easy. it's really hard to take away something that users love in our familiar with to lead them down potentially, and that's a big keyword, a better path. we aren't right one hundred percent of the time, but there are two that jumpp to mine. when i was putting together of the slides with this talk. the first one is giving ai a terminal. we all remember fears about son of anton deleing your entire code base and doing catastropc things to both. you know, you're start up your company exetra exetra. but as models got better and people invested in primitives, such as permission systems, users ended up building, faster ended up shipping more. and they did so safely. so we're able to overcome this. and as models got smarter, they're able to bake better decisions about what they should and should not run in your terminal.

Speaker 4: the second instance is this tweet, which is very representative of sort of the yelling that i got when we removed chat from wininerf. so a lot of users were yelling at our team because we took away something that was very dear to them. the chat side bar and replaced it with only an agent. now at the time. this is something that was familiar. and rather difficult to swallow. but when we look back models have advanced multi step research, agentic research and execution became the new paradigm. and here we are today using and loving all these agentic products. and so now i bring you to today's battle, what is going on today. so we did couple the agent manager from the id. and with antigravity two point, no, we spot them into separate applications. we believe that the ide is to the agent manager, what the debuggger was to the ie. you don't always need a debugger, but it definitely is helpful to have it. if you need to go a layer beneath and go one step deeper into that abstraction stack. and our prediction is that this idea of agent orchestration, you can call agent teams. you can call it swarms, you can call software factories is the future, and we're willing to bet on that future.

Speaker 4: so here are the primitives for what we're calling. the agent teams twenty twenty six era. these are things like subagents generative UI inside cars. we'll talk more completely tly about those things are in some examples of how they manifest inside of the product. but it's really important to first understand the why what brought about these changes and what model changes, what model properties actually led to the development of these new things? and as a product team, do you force the new era of primitives? or is it something that comes to you by using the model and experiencing the model? the answer is kind of both right and the privilege of being inside of google deep mind is that we do have that relationship between the product and the model. so you remember the crox of antii gravity, one point o is to manage agents in parallel to put the human in the driver seat. and if you remember my last talk, i talk a lot more about this research, product fly wheel. and now as promised because of the antigravity product, geri has now learned a finger, too about how to manage a team of agents. there's still a lot of headroom room, make mumulti ents systems better, more collaborative, better, a deconstructing task into smaller task. but we've got a really good head start with germany. and all the basics have been imbuted the model so that we can build up product like antigravity tupoint out gernine three point five flly gmini three point five flash laununed back in april. and this brought to market a lot of those capabilities that we have been working on on the background with antigravity and flash now isn't just good at executing test. it's actually really good at leading teams s. it's faster and cheaper, pushing the parado curve of what is intelligent, verses, the speed and the cost at which you run those things. and putting this all together, we were really excited to announce agent teams in public preview inside of antigravity. all you have to do is simply type the slash command, sash teamwork, and you'll see a new mode where you can enter an unleash, a swarm of agents onto the task at hand.

Speaker 4: so we'll talk a little bit about how this works. you as a user will specify your task. the more specific you are the better, though, the nature of these agentic communication styles is that if it need something more, it can actually ask you for more until everything is basically clear. you'll work with that lead agent, and it will manage a team of arbitrary size to get that work done. and when i like to say it's kind of like the advengers right didon't take a bunch of specialized roles. it made front end engineers back in engineers, infrastructure specialist QA design, the list goes on and and on and on. there, there are infinite possibilities for what each of those subagents could take on, each sovasion is dynnamicgenerated and can operate independently, and it can even actually select a different model from what the main agent is using. and this is done so by that main agent, again, we are scaling with intelligence. and one of the coost aspects of this is that it can use generative UI with a model that is as fast as flash. things can happen nearly instantaneously. if you ask, hey, what is the status of my task show me a camin in what's going on, or maybe you know, prefer something a little bit more like, uh, the chrome debugger tool, it can show you a timmlike. and all these things are generated on the fly because it's able to generate UI on demand.

Speaker 4: so some of the projects that the system has implemented. we've built a photo editor, you can actually edit raw photos directly inside of your broser. um we've also built a messaging app that might look a little bit familiar to those in the um in each each. these took hundrereds subagents uh and took almost half a day to run, but to really put it through its paces. one of the hero runs that we did was actually building an entire OS kernel. this is something that weve got to show off at google io, but we built a complete os corneal from scratch and actually played dooe on it. and my colleague room was able to demo this at google io. we were superproud of this particular milestone because it really demonstrated that if you throw more intelligence, you threw more subagents um at this sort of problem a model like german at three point five flash could do this in a way that was not only very, very powerful, but also scalable. and, you know, mild, dly, affordable. obviously, we're not going to spend thousands and thousands of dollars to about an OS correnal every day, though it is possible. and some of the stats out of this, it took ninety three subagients over the course of twelve hours, made fifteen thousand requests, two billion tokens. and it was under a thousand dollars, which was one of the really co aspects of this project. and so as you can see with this particular example, subagent primitives are one of the defining parts about building um twenty twenty six era of agent teams.

Speaker 4: so agent teams are just that first example, and i want to show you another example that our team uses interternally that sort dedestrates of these new primitives. um the second one about auautomating research tasks, so we work inside of gemini. we help sort of make gmini better at coding related task, agentic related task. and this is where the real magic starts happening with the product. we have an internal version of antigravity that researchers, engineers non technical folks can use. and when they understand the primitives that antigravity offers, it becomes a very, very powerful way to automate your own worklose.

Speaker 4: so we'll take the example of side by side eval analysis. this is a very common workflow, not only a deep mind, but just generally in the industry. you essentially, you will take multiple rollouts, one, two, three, four extrtra, and you want to compare them. so you'll take a set of test. you'll do some rollouts youll get some results and y'll essentially be in two different tables. 对对, now you'll look at contcontrol, look at the experiment ment and know how to figure out not only what the difference was. but perhaps what are the reasons for those differences? and how can we actually iterate from there and make a better version of for the next experiment? now? traditionally, this was a lot of jupiter note book elbow, greece essentially. but when you start working with the new primitives in twenty twenty six, you end up with a lot cleaner of a worklobe.

Speaker 4: 嗯, so researchers wewere able to automate ninety percent of this workflow by simply asking the agent about the evalils in questioning using natural language, then the agent that is now primed with skills and an understanding of googles. massive montreo code base is able to crunch the numbers and get back to you with a delta. now what's really cool here is instead of just taking that delta than handing it to the user, it went the extra step. it spn up for a research agent specialists that proposes a hundred different exhrothesis over why those deltas might occur. and then it uses subagents to then spit up one subagent for hypothesis and basically drills into that particular case in parallel, mapping back to a single response and then telling the research are. hey, here are some areas that i found. now, here's a report that you can review.

Speaker 4: and what's really cool is that it doesn't stop it, just the report. it actually puts together a generative UI for you to look through interact. so i drop down filter segment slice and actually interact richly with that data. 嗯, and internally, we care a lot about this sort of workflow, improving the model, improving the product and understanding the ways that users find success and failure internally a google. so what used to be a very manual process. now it takes minutes. so what used to be hand engineering, you'd have to build your own acing pool of agents, you'd have to set up your judges, you'd have to take together data pipelines. all of this now starts becoming grounded in these new primitives that we established earlier in the slideshow. you have a subagiing graph that is completely dynamic. the generative UI comes in the end to rithally convey the findings in a way that the user best understands, or maybe vocaters to their learning style. and all of these things can be regenerated and redone on the fly. all the user had to do was load up a skills file and asked away 没有异异议.

Speaker 4: so with teamwork and this ebel example, we start arriving at these twenty twenty six primitives that i keep talking about. and these model characterteristics really changed the way that we have to think about the product and how we have to develop the product. so the three examples we've talked about, first, we have the dynamic ovagent and to provide a little bit more color here. 嗯嗯, basically, no two subagents are the same. the main agent is the one that is orchestrating this entirely on its own. it's configuring and prompting and ceding the subagents on the fly. they they ooperate in parall l they ooperate in different types of secure environments be at a sandbox be at a remote execution system. and they can all take on ininfinite number of specialized roles. so the scaling story here is quite obvious. and from the last two examples, you can probably tell, as the model get smarter, your team will become more specialized. it will become more collaborative. and ultimately, that means it'll be capable of getting more complex work done for you.

Speaker 4: and now the second is this new concept, we've uto do it slightly in the past, but it held side cars. this is a new plugging protocol that we're bringing to integravity a side car ar prois esessenally ally. it is a decar process. the naming sort of reflects what what's going on to the the hood, but it's a long live utility, and it's responsible for listening. it allows the model to listen to the outside world and set up its own triggers for things that might happen. for example, this could be sms messages. this could be web hooks, cron jobs, hooking it up to get hub PS. the list goes on and on, but this is a generic plugged in primitive integravity ity already use a side cars for things that our time based. this is where the the scheduled test croon concept comes from. um but under the hood, this is all this new side car primitive. so when 're releasing the spect for this, so that you all can build on top of this new primitive um later this summer, but there are some really, really cool ways that people internally have been using this sort of concept.

Speaker 4: and the third and final primitive is generative. i so we hypothesize that human written specialized us are kind of dead geri flash on antigravity clocks in it almost nine hundred tokens a second. this is ten x passive than a lot, the other frontier model experiences. and in a matter of seconds, you're able to go from whatever you are thinking inside of your head into a prompt into a use case that is designed and embedded inside of your conversation view perfectly. and rather than rely on tempplts or even even t ml files antigravity can render your generated UI in line, so you can do things like things in play doome. but this also extends the things like bar charts, graph tables, anything that you would want to interact with and maybe inspect a little bit further than just a markdown file or just a conversation in genertive view in, in many ways, reminds me of the quote that the late steve jobs said, when unveiling the iphone, he justifies the removal of the keyboard. and says they all have these keyboards. they are there whether you need them or not. and they all of these control buttons that are fixed in plastic and are the same for every application. and in an nawgious way, we built our product to dynamically scale with the needs of the agent, we skipped to have a infrastructure and mechanical us in favor of sidecars in generau. i and that creates a product experience that is not fixed in plastic.

Speaker 4: so subagents side part triggers in generative UI are the latest permitives that are powering antigravity. we've tried our best to stay out of the way and let the model cook. and if you're building a product around an agent, you should consider what are the primitives that are in my product? and how might they scale with the model's intelligence? we all are familiar with shipping features is now quite easy with all of these new tools. and it's about deciding what features to actually add so that the model. so your product can scale with the next release of the next model, which will inevitably be faster, better and cheaper. and so with the right primitives, you as a builder or you as a product owner, you might be surprised of what the models can do.

Speaker 4: and in classic fashion, i'm going to keep using the slide until i've actually conquered the TPU crunch. so you can find me on twitter. uh, you could dm me for feedback. we're always looking for new ideas on how to build the latest sting greatest. thank you for watching. thank you, swix and bent for having me. it's always joy to be here, and i'll be at the antigravity booth if you want to talk further. so you want to get to know the product a bit more. some t members will be there. so thank you so much for your time,

Speaker 6: if cited to meet you all not for human in twenty twenty six.

Speaker 11: coding agenents will quietly retire the first soft platform, not because it's bad simply because the platform is unnecessary. i am dominic tunnel. i am founder in ceo of resonate. resonate is a global execution platform built with minimualism and simplicity as its core technical values, and these properties will play a central role. in this talk, it resonate. we have a working theory where software engineering is headed general purpose implementations will increasingly be replaced by be spoke implementations. generated on demand not is a new library, a new framework or a new platform, but is a minimum extension of the infrastructure that is already in place. this theory holds too reuse will move upstring instead of reusing, a general purpose implementation. we will reuse a specification, and we will derive a be spoke implementation from it. in fact,

Speaker 11: we can build many bespoke implementations tailor made for the infrastructure that is already in place. We just have to ask the agent at this point. The prompt is a platform. Resonate is a global execution platform. We have an implementation of the Resonate Java. We have implementations of the Resonate SDK for TypeScript, Python, Rust, Go and Java. So we have to ask, what does this new reality mean for us if implementations become generatable, where does our value live? And our answer: our value moves from implementation to specification. Now this changes how we think about reasoning, the product is no longer the implementation. The product is specification, the protocol. And from that protocol, we want to derive multiple server implementations. One is a general purpose Resonate server, our reference implementation, others are implementations built with infrastructure partners for customers and partners. This means durable execution right on top of their existing infrastructure with minimal additional dependencies. So the question is no longer can we build a server? The question is, can we repeatedly synthesize trusted servers from the same specification? And if so,

Speaker 9: how when we talk about agentic engineering,

Speaker 11: we focus all of our attention on verification. How do we know the result is correct? But today, I want to focus on the specification instead. And more importantly, how can agents participate in specifying the system, not just building or verifying it? Now recently we partnered with multiple infrastructure providers to bring durable execution natively to their technology stack. One of them is Syndia, the company behind NATS, an open source messaging system designed for building modern distributed systems. For the rest of this presentation.

Speaker 0: we okay.

Speaker 5: hello, everyone. I'm excited to be here. My name is Zac Lloyd. Today I'm going to be talking about self improving software factories, the new open source model. And basically, what I think is happening to development. A little bit about me just to begin. So I am a former principal engineer from Google. I led engineering on the Google Docs. I've been an engineer now for over twenty years, a long time. I'm not shipping frequently, but I haven't read Atlantic code in the last six months. And I'm the founder of a company called Work. If you're not familiar, it's an open source agentic development environment. You may know it as a terminal, that is how the company started. We were basically a terminal that has agents built in. We open sourced it a couple months ago, and I'm going to talk a little bit about that experience and the motivation for it. It's a popular open source project, over sixty thousand GitHub stars. We've had a couple hundred people contributing, over eight hundred thousand active developers who are using Work. But increasingly, we are focused not just on the terminal aspect and the interactive aspect of developers, but more so on how do you automate development. And we'll talk mostly about that. So the thesis that I have is that the discipline of software engineering is going to become something more like factory engineering. And I'll explain what I mean by this in a minute. But just keep that in mind, that's what I think is going to happen.

Speaker 5: If you look at development over the past decade, it's crazy how it's changed. We've gone from a world of chat and auto completes, Cursor or Copilot, to the phase we are in now, which I consider to be mostly interactive agents. So you're sort of sitting at your computer, and you are telling Claude Code to do something, you telling Work to do something. And I believe what's going to happen over the next six months, a year, hard to predict, is that we're going to move much more towards the world of automation. But before I dive into that, just a quick show of hands: how many folks here are building agents? Almost everyone, makes sense. How many folks are building typically with multiple agents at one time? Again, almost everyone. How many people are running an agent right now? I am not offended. I would get the ticket, I'd be doing it. So how many folks are running agents in the cloud, out of curiosity? That looks like less than half, but still significant. And how many folks have set up a system internally to automate the whole of the software lifecycle? So everything from like triaging, specifying, implementing, reviewing... some hands. So some people are doing this. So this is what's going to happen. Every project of significant size, I believe, is going to have something like this, and it's going to look kind of like this big loop. And everyone is talking about loops. There's nothing that complicated about loops. This loop says a cloud software factory. This loop could literally just say the software development life cycle. It's the same thing. But just to go through this loop: ideas are going to come in at the top, agents are going to do triage. If something is complicated, they'll spec — these old blue boxes are where humans come in, humans will review the spec. Agents will do the implementation, a human and agent will review the code. Agents will verify, human will review the product you ship, and then you monitor and round and round you go. And this is what software development, for better or worse, I think is going to end up looking like.

Speaker 5: So I repeat the thesis, which is that if this is what software engineering is going to look like, software engineers are going to be the ones who end up designing and managing these factories. Now I promised at the beginning, and I put it in the title of the talk, that I would talk about open source. And so I want to do that for a few minutes. I bring up open source because one of the main reasons that we open sourced was to build a public factory. And so this is a picture of this website we built called build.work, which shows all of the issues that are flowing through our system and what state they're in and what agents are working on them, what contributors are working on them. And it's kind of like a proto-factory done at scale. It's not working perfectly, but it is working. And one of the reasons we open sourced was to try to build this. Just in general, I think it's interesting to talk about open source in the time of agentic development. This is a really super graph, but it's like, you get it. It's like it's becoming much cheaper to build software. A corollary of that is that it's becoming trivial to clone software. And so if you are in the software business — I don't know how many folks in this room are in the software business per se — but it's very hard to build a software business if it's free to build software. It's hard to capture the value, especially if a competitor can clone. And so my big takeaway, or big tip, for everyone in here is: the first thing you should do is patent your code. I'm kidding, don't, this is a complete joke. Don't do this. My first step is obviously, you need to have a great product. This has always been the case, but I would say a great product probably was never enough, but even now more than ever, if you think that you're going to build a great business just by building and shipping a great product, you're probably not gonna succeed. You need advantages beyond the product. And so those advantages could look like distribution, ecosystem, could be that you have a great brand or data moat, you might have capital. But if you're a startup again — I'm coming from the startup world here — you just don't have these advantages. And so you still want to break through. And one of the ways that I suggest doing this is by building in the open.

Speaker 5: So to be clear, it took Work five years of building closed source to sort of make the product, and then we decided to build in the open and explain why. If you build in the open, it helps build your ecosystem. It can take you from like hated on Hacker News to like tolerated. It can burnish your brand, create community. And so there's all these advantages. And some of the things that have traditionally been a pain can now be managed. And so like the traditional pain of open source might be something like you get a lot of noisy issues, you get sloppy PRs, you can end up in code review hell, you can end up having to spend a lot of time verifying changes. And so the solution — kind of a long way of getting to this — for open source, or at least for Work, the thing that made us finally decide to do this was that we built a whole set of automations, really a software factory around managing the open source project. And so like I said, this is what I think the future is going to look like. I'm going to dive in a bit, get a little more technical for folks who want to try to build something like this for their own projects.

Speaker 2: Mm.

Speaker 5: So what are the components of an effective software factory?

Speaker 5: it's really not that complicated to start or at a high level. you need a set of automations. you need a way of providing context and skills. you need a way of bring humans in at the correct time to sort of like when things get stuck on the factory. and then a really important thing as you need some set of self improvement capabilities. so think of this is loops. and if you do this right in the open source world, you can get a, you know world where agents are helping contributors contribute, they're helping maintainers maintain. and one an emphasize, there's nothing special about open source here. i think every sizable project can benefit from this approach. and i predict that every every company, every open source project will have at its core a software factory kind of lect the way that CICD became just like, of course, you have that. maybe i don't know when that happened ten years ago.

Speaker 5: let's tour the attack the factor floor for a second year. so you're not going to look at this. this is too much. um the point of this slide is not to have you read the worflolow, it's that facfactory for or is basically a graph um of steps where you are defining like, okay, how to to offer get built for my product. and it looks pretty similar for every product. things come in. they flow through um they get stuck at certain points um and you know, broadly speaking, just to back out of second. so there's the inputs, the inputs are really ideas. um the inputs could be coming from your team. they could be coming from your users. um the inputs themselves can end come in through certain channels that you should think about, as like your cass trackers and obvious one or slack, your teams, like your communication channels. they could come directly from like your terminal RDE. they could come from your monitoring systems, but there's some set of inputs that bring work into the factory. there's trioge. this is a really important step. so again, i boiled down a sathing, very simple, but like you want an agent that is looking at issues as they come in into saying, you know, if this is easy, and this is unambiguous, just simplement it. and this is how you can actually get going with a factory. if finissuyou is hard. if i recommend having an agent that produces spespects.

Speaker 5: you can do this many different ways, but i think it's very effective the way that we do it at work that i recommend is having an agent, right? what we call a products spect and the tespeproduct spect describes the product and variance the are building towards text, back describthe architecture and the shape of the code. then you have an implementation agent. this is basically a coding agent that run somewhere in the cloud. it makes a def. you can use all sorts of coding agent for this. you have review this isn't many ways. the most painful part, like i expect, the people are a little bit tired of reviewing identical slop. i would have an agent two code review first, and then it becomes over time, like a risk management exercise of like when you bring in humans to do coter view, but you want to have a step in heir where humans can do it. this is a very important step for certain types of apps. the verification step. so this would be things like computer use you're building AA sort of UI having the computer actually used the code at the agent, produced and producing videos and screen shots. 对对对, CICD still use it, obviously. and in monitoring. so agents don't stop in your factory when code a shipped, they should observe what's been shipped? is it crashing? is it being used and round around you go because you take the output of this monitoring step? and you feeit back into the top of the factory.

Speaker 5: now you can try and build this. and i went to a talk earlier that my friend anamm gave where uber has built an internal version of this. um and it's pretty cool. i would say, for most, you know, most organizations, it really depends where you are. you'll be able to build a simple version of this easily, but to build a thing that actually scales is probably like you's, probably focusing on your own product, not building this infrastructure because there's a lot of stuff that you end up wanting. you're not think king, not what to read this. it's just a lot of stuff because you do build it. um if you you it, it, you ll end up with something that looks kind of like this, which is um you're going to have a bunch of ways to getting work in the your factory. that's what's in the topiyear. you're going to have a sort of control ane for figuring out how work gets distributed across your factory floor. you're going to have the actual place where where the work happens. and so that's going to be cloud sandbox is it's going to be figuring out what agent trun. so what's the harness? what's the model? and then finally, i think this is a really important thing. you're going to want to set up some kind of data plane that sits below your factory. and so that's something that lets agents remember what theyve done, learn, improve overtime.

Speaker 5: the factory is not just like a product. it's also a mindset. and this brings me back to the thesis i had at the beginning, you need a measure and improve. so factory is is where they i don't know. you can stretch this metaphor as far as you want, but like you should be thinking of efficiency. and so that means like how much software did you ship, how much should it cost in terms of human time and token time? and you're going to want to measure this and try to improve it over time. a key part of this is creating loops. so loops are again, they they sound complicated. they're not that complicated loops are basically ways of having agents and improum by like observing what they're doing, where they're failing. um and so a common kind of looops you're going to want to in your factory is like a skill loop. that means you're going to have your factory agents that are running skills, and they going to have observer agents that are seeing how those skills are being applied um for it shoes and trying to improve the skills. so for instance, if you had a code view view ent um and it was leaving comments, and a singer engineengineering, your team is going correcting. those comments you'd want an observer agent that would look at that and basically um improve the coter view agent for the next run.

Speaker 5: this is one thought just to leave folks with, like, what is where is this live engineers? i think you're going to have to get into this mindset.

Speaker 0: and i'm trying very hard to get our team into this mindset. it's not always easy that you're not just building the product, but you're building the thing that builds the product. and that's like it's just different. it's more like process engineering your manufactureing or something like that. and you could think like, okay, maybe that's obbuomer like is that obomber is that you know? and it depends like it depends what joy. you get out a software n engineering if your joy is in writing the code. i think you're everyone in here who is a software engineer is going to be writing less code. but if your joy is in shipping product like it's never been a better time. and this is actually where i find my joy. it's like i like building in shipping the thing. so everyone in here is going to code less, but they're going to shiit more and that's going to be a trade off. but if you approach it like you're a factory engineer, i think you can see like that they're still a really cool set of engineering challenges. you could almost think it was like matter engineering, like, how do you engineer your system of agents to be the best possibility engineering? but i think is a very compelling and interesting set of challenges to solve.

Speaker 0: so that's it. uh for folks who are interested um this. if you follow the income as QR code, i've set up um a open source skittobrepo, where anyone who wants to try building their own factory agents can do it. this uses warps uh agent platform as part of it. but you honesay you don't have to use it, but i'm not trying to push ing into our product, but this should give you a good sense of like, okay, you want to set up uh, an agent that does um trios or an agent that does spectwriting. how do you actually do that? how do you get from like the theory of working with a factory too actually putting it into practice? um i don't know, we have the capability do questions in here. i saved a few minutes for questions. if anyone has questions, otherwise, i will. i wrap up, yes, yes ah, you want want to build this. it's a lot of more. yes, you are also a it's a great question. so so said some of it alalconcontraicictory, i think you should the way you should think of it is like everyone's going to deploy some sort of factory. but then the tuning of the factory, though, like are these the right skills for my domain,

Speaker 0: Is this factory building my product in the right way? I still think there's a bunch of interesting engineering challenges. And for some places, you can build this, but I get I think, think, think, think that you, you should probably be focusing on building the core product for your company for the most part. But there's a bunch of like tuning and like figuring out how to make the factory work for your product, right? That matters is a great question. Yes. Is that thank you so much. I had a question. So if you were a college student right now, yes, graduating and entering the workforce where would you spend your time yet? So the question in case people couldn't hear was like if I was a college student graduating and entering the workforce right now. So I think that the most important skills in this new world are adaptability. I think that's critical thinking. It's like the ability which you can learn. I do think I don't know if you're a computer science student, but I still think there's a ton of value in understanding like the underlying systems and architecture and being able to reason and understanding, I think, could understand the specs that agents have written. I started writing. So I would focus on those skills. And like, I don't know, we're hiring more people than we've ever hired. There's a lot of kind of like misdirection around like, you know, people not being hired because of AI, that's not our experience. So far and what I'm looking for are like really adaptable product focused thinkers who can make basically any problem, solve problem solvers even as the underlying technology changes. Yeah, I think of time, for one more question is what about the like? So the question was, what about the product taste? And like how do you actually build useful? I think that's probably right, like maybe the framing. And like where do the ideas come from? So I think that um there's a problem with the, the factory metaphor, and I'm like leaning into it because I think that's that. There's something to it is that it can kind of sound like um mechanizing or dehumanizing. Um and I still think underlying all of this. The only thing that matters is like, are you building something useful? And if you have like a factory that is like turning out stuff that no one cares about. It's like what's the point? And I think that human taste, human input, human product sense, humans like guiding at those touch points where you can't automate stuff is that absolutely like essential. And like that's like what I do. I'm trying to figure out what what do customers want, what do people want? What's can be valuable to them. So I think that's absolutely key point to it. I think I'm out of time, so I have to go. I hope folks enjoyed this chat really grateful for being invited to speak.

Speaker 2: So thank you all very much. Agents in production.

Speaker 4: Specifically how OpenGov built and scaled OG Assist. So um this presentation is going to be jam packed with just so much good stuff where we talk about AI agents, we're going to talk about our harness, we're going to talk about um evals, observability, traces, we're going to talk about um tools and skills. Um it's there's going to be a lot of good stuff in here. We're going to talk to you guys about what we do at OpenGov and how we operate at the scale that um we operate at um in production. So you'll be able to see a real use case and workload uh with AI agents. Um so without further ado, let's get started. Okay, agenda. So just really quickly going to go through um high level. What we're going to talk about today. Ah I'm going to tell you guys a little bit about OG Assist. I'm uh OpenGov is I'm going to tell you guys the origin story of how this all kind came to be. Ah we're going to talk about OG Assist, uh big bet on Effect, uh, a little bit into our agent loop, we're going to talk about the A2A protocol, evals and sandboxing. We're going to talk about how we manage long context. We're going to talk about um monitoring, observability and how we collect feedback um and how we iterate on that feedback. We're going to, lastly, uh also talk about tools and skills and how at OpenGov um we use um AI not only externally that we uh serve to customers, but also internally to improve our development workflows. Just a little bit about me before you go any further. My name is Gabe, I'm a software engineer here at OpenGov. I work on the AI agents team, and I'm one of the folks that helped build um OG Assist and some of the systems that you guys will be seeing today. So a little bit about OpenGov. OpenGov is a software company um on a mission to power more effective and economic government. Um so OpenGov sells ERP software for things like budgeting, procurement, asset management and permitting. And um we were founded about four years ago. And what's cool is um we have this thing called OG Assist and OG Assist is this little button on the top of all of our products in the navigation bar. And what's cool is um all of our product suites and product teams um have built tools and skills in order to power this button. So for example, if I open up um this, um if I click this button, I opened up OG Assist says, hey, um I'm going to ask about Racos, which is very specific to the current product that I'm in, and you can see that inside of this kind of chat interface, I'm able to speak to an agent. And the agent is able to make tool calls in order to um look up information against data inside of that suite. So it's really cool um to be able to kind of first party create these experiences through the capability that we built called OG Assist.

Speaker 1: Right.

Speaker 4: Okay. So just a quick story about how this all came to be. So um a little while back, we saw that AI was really starting to take off in a principled way, the AI enablement team asked me to join. And um instantly, I said, yes, and OG Assist started to grow. And we started integrating OG Assist system into our products and not only our backend capabilities, but also our frontend capabilities as well. So you'll see that one of the capabilities that we give the agent is to um see what's on the screen and and take action on what's on the page. So you could see that um I'm asking the agent here, hey, what's on the screen? Can you maybe highlight um some the next steps that I could take? So you can see that the agent here is thinking, it's saying, okay, what tools do I have available to use. And hey, let me go in and highlight something that you could actually click on and tell you more about it. So just a capability of the OG Assist system. Just a little short story about how this all came to be. So the big bet on Effect. Um so I really wanted to include this slide because um here on the agent team, we made a huge bet to bet on Effect and suffice to say it's paid off in dividends. Um we write Effect. So Effect is this library for TypeScript. It's open source, and it helps you write better TypeScript code. Um you know, it's got a lot of stuff like Schema, like Expo. If you've ever used that, it's also got um things for error handling, for logging, for traces, it's just got so much in there. It really helps write better code and structure your code better. And it helps with architecture and spinning up new services. And for us on the agent team, it really helped design and build the core agent loop. So you'll see throughout this presentation sprinkled in um how Effect on our team uh has paid off in dividends. So we really love Effect here at OpenGov, and we encourage other folks to try it out. Um yeah, let's keep going.

Speaker 4: The Effect native agent loop. So originally, we were on LangGraph. And that was fine until the team really started to scale, and our use cases started to evolve. So we decided to move over to our own kind of Effect native agent loop to have full agency over this agent loop, such that when we have complex use cases or features that we need to build, we could kind of get in. We had full control of the agent loop, but not only that, but now we're fully on Effect. So all the cool things you get with Effect now are propagated throughout the entire agent loop like the tracing, structured concurrency, logging. Everything is more fine-grained control, and it really allows us to really unlock the full potential of having our own agent loop from the ground up.

Speaker 5: Uh.

Speaker 4: So another thing I wanted to mention is on the left you'll see a code example. This is really the basics of the Effect loop that we're using. We're using this thing called the Effect AI package. And in that package, there's this thing called um there's a Chat and a LanguageModel. So with the Chat, you can instantiate an AI chat, for example, and then you could stream text using um that kind of streamText function, you could pass in a prompt and it'll work with a language model under the hood. Since we're kind of doing dependency injection, we could pass in a different language model if we were to hot swap to another one.

Speaker 4: for example.

Speaker 4: so really just having full roll of our own agent loop,

Speaker 4: just kind of gives us all the levers.

Speaker 4: and it really just unlocks the full capability of the model and for the team as well to have full agency over this.

Speaker 4: this loop.

Speaker 4: no,

Speaker 4: the thing i wanted to mention is the agent to agent protocol.

Speaker 4: so here on the agency,

Speaker 4: we've had a lot of success with this protocol.

Speaker 4: so this protocol being the protocol that google created kind of an open protocol for agents that in our communicate.

Speaker 4: but um we found this very useful for um defining our agent rourouts,

Speaker 4: for example,

Speaker 4: in the back end and our model and our schema to follow this kind of a agent protocol.

Speaker 4: so we model.

Speaker 4: so for example,

Speaker 4: there's this thing called an agent card,

Speaker 4: what you see here,

Speaker 4: and it's got the name of the agent,

Speaker 4: a description exetra,

Speaker 4: right?

Speaker 4: and having this kind of a rigorous protocol.

Speaker 4: this rigorous spect really helped drive our development and drive alignment,

Speaker 4: because you know,

Speaker 4: all we had to do was um gying with this spect and folall.

Speaker 4: this spect.

Speaker 4: and we knew that this was kind of the contract that our front and and back in with both consume and and produce.

Speaker 4: so um this uh,

Speaker 4: i would say also has been um very helpful for us.

Speaker 4: and and what's really cool is a to a has a lot of extensions,

Speaker 4: right?

Speaker 4: so you can extend the protocol um adding like meda data.

Speaker 4: um there's also a to UI.

Speaker 4: um so lots of fun stuff,

Speaker 4: uh with a to a protocol,

Speaker 4: uh h is kind of what's worked for us.

Speaker 4: so to sharing that with with you folks,

Speaker 4: feedback and evots.

Speaker 4: so here the quote is shipping is the start,

Speaker 4: not the finish.

Speaker 4: so what we do here on the agenscene is we have kind of multiple ways we do evalils and collect feedum.

Speaker 4: obviously,

Speaker 4: you know,

Speaker 4: will have foks s call in or email us or or just let us know and tell us.

Speaker 4: but the main way is we have the stumbs up and dumbs down mechanism.

Speaker 4: and here um someone is able to tell us,

Speaker 4: hey this.

Speaker 4: this work really well.

Speaker 4: this was a great response,

Speaker 4: or that wasn't a great response.

Speaker 4: and that signal,

Speaker 4: we taken reable to iterate on um and we could take it back and help and prove um you know,

Speaker 4: the response in the future.

Speaker 4: um we also have automated evalils.

Speaker 4: so in in the in our CI,

Speaker 4: we we have evalils that run against real completion.

Speaker 4: so we could test the prompt against.

Speaker 4: hey,

Speaker 4: did it hit some tools did to do what it's supposed to do.

Speaker 4: and that also helps with our accuracy.

Speaker 4: so those automated eevalls and conunction with collecting feedback really help.

Speaker 4: and so improve are our our ools ls.

Speaker 4: our skills are harness ness,

Speaker 4: and that's really how i were able to iterate so fast.

Speaker 4: and so quickly,

Speaker 4: humans in the look.

Speaker 4: so this is really coofeature ure built t where we demministically in our up the agent loop,

Speaker 4: if there is a tool call approval required.

Speaker 4: so if in agent ttrists,

Speaker 7: okay,

Speaker 7: hi everyone.

Speaker 7: i'm in a sotomon.

Speaker 7: i am the creator of eighent craft,

Speaker 7: and i also created MCPYI,

Speaker 7: which became MCP apps and MCCP steering committee.

Speaker 7: and then a bunch of our stuff.

Speaker 7: but we hear to talk about why roll the ball on neck and why we don't have to be.

Speaker 7: so uh as you all know,

Speaker 7: and user b day agents are amazing.

Speaker 7: there's so much they can do from generating cat memes to b to b saataps.

Speaker 7: so if there are so amazing,

Speaker 7: we have to ask,

Speaker 7: why aren't we all unstppable?

Speaker 7: why doesn't everyone have an army of agents that does whatever they want,

Speaker 7: so it should be relativezy.

Speaker 7: i mean,

Speaker 7: you just have a bunch of modeitorors,

Speaker 7: and you spin up twenty five clouds code engines.

Speaker 7: and you're done.

Speaker 7: so the reality is that it's just not that simple to do something like that,

Speaker 7: you need to a really want to use agents.

Speaker 7: and secondly,

Speaker 7: once you actually spend them up each one,

Speaker 7: these these agents requires you to steer them,

Speaker 7: direct them,

Speaker 7: review them.

Speaker 7: and when it's a scale officially,

Speaker 7: it's just exhausting.

Speaker 7: we need to find the scalls to allow us to do it without burning out.

Speaker 7: so in reality,

Speaker 7: we are the ballonic.

Speaker 7: and the cool thing here is that why we are the ballonic,

Speaker 7: the skills that we need to do these things to find ourselves,

Speaker 7: not burning out every time you used agent or will ver.

Speaker 7: so along ah if you try to comare it to our stuff that you've been doing outside of programming,

Speaker 7: for example,

Speaker 7: playing video games.

Speaker 7: it's not really that different,

Speaker 7: like if you're playing workraft,

Speaker 7: for example,

Speaker 7: or seems,

Speaker 7: and you have a bunch of different units running around and need to understand what you're doing and supervise all of it.

Speaker 7: how definn't do that from from actually inspecting agents.

Speaker 7: so i took this concept too far,

Speaker 7: and i created agencraft,

Speaker 7: which is a game inspired orchestrator and tries to bring all that good stuff together into work.

Speaker 7: so let's take a quick look of what that actually looks like,

Speaker 7: can try and understand the journey that we can go through to really raise the feeling of how humans and agents can collaborate.

Speaker 7: so this is a agencraft.

Speaker 7: there's really a lot to unpack here.

Speaker 7: so we'll just go for the basics to a quick what to see here is an agents.

Speaker 7: um it's an actual representation of a cloud code or open clode.

Speaker 7: whatever agent did,

Speaker 7: you can imagine it can be detected on your device,

Speaker 7: or you can actually spond new ones directly from agent craft.

Speaker 7: um so nex said you have like codex and uh open code to setrum.

Speaker 1: 了解我的在的 at your command,

Speaker 7: and these agents can do everything.

Speaker 7: they're just normal agents.

Speaker 7: so i can just prompt them in this nice siypanit.

Speaker 7: it also has voice.

Speaker 7: so there's moking maddhere uh and just do something for me,

Speaker 4: prompt prompt.

Speaker 7: and now these agents can do whatever.

Speaker 7: and if you look around,

Speaker 7: you would see that the map has a bunch of different eleelements.

Speaker 7: so so you have buildings.

Speaker 7: these builings reprepresent fununctionalities for for exple.

Speaker 7: if if you can manyour your lulugging ssls um you have like integrated terminal integrated get.

Speaker 7: you can can do your all of your work from within this space.

Speaker 1: 对对对对,

Speaker 7: so not as we have this and we understand the basics,

Speaker 7: let's look at what we can do and to raise the scening.

Speaker 7: so the first step is obviously,

Speaker 7: visibility.

Speaker 7: i neeneed to understand what's going on.

Speaker 7: what are my agents doing?

Speaker 7: so i have this my cpphal,

Speaker 7: and it shows me the different agents and what your task is and what the left me they did is and what their currencdoing everything i need to quickly understand who needs my attention and who doesn't 对对对,

Speaker 7: the second part of it is when not just limited to this,

Speaker 7: you have a map.

Speaker 7: so i can take that,

Speaker 7: say my file system project it onto the map,

Speaker 7: and i can see exactly what my agents are doing,

Speaker 7: not just by be looking at the list of stuff,

Speaker 7: i can actually see.

Speaker 7: okay,

Speaker 7: this agent is not working on this directly or this file and the ile ile that presented as rooms on the map.

Speaker 7: so i can really look at the tron and see exactly what each agent did and when and i can take this nice information because we are do aucus later and use that to create lingages.

Speaker 7: and i can create cold stuff like heat maps.

Speaker 7: so i can really easily see what's going on.

Speaker 7: and 对嗯 the the the other aspect is that now does we have this visibility?

Speaker 7: that's not really enough.

Speaker 7: we need to be able to react to what's going on.

Speaker 7: if one needs my my attention.

Speaker 7: i need to be able to quickly jump to it so much like video games or RTS games or civilization,

Speaker 7: whatever you would click space bar and would just take you to the nearest thing that actually need your attention so you can really easily understand is going on.

Speaker 7: and jump from one thing to the other and answer questions or true plans.

Speaker 1: and you're jorun that was nice.

Speaker 7: i meit's a nice step just to have that disability,

Speaker 7: but it's not pretty enough.

Speaker 7: it's pretty exhausting to keep looking at what the agents are doing in the grounllarity of what father are using and what tools are using at the moment.

Speaker 7: so we need something else.

Speaker 7: and the reason is that we have,

Speaker 7: like two main problems here,

Speaker 7: one is that there's this too much to hold in our heads.

Speaker 7: if it,

Speaker 7: if i can understand what's going on,

Speaker 7: and i can see everything i can't pretty like twenty things.

Speaker 7: at the same time,

Speaker 7: i can't even think of the next thing i wanto do if you have this huge coding project to whatever.

Speaker 7: so i need to find some way to fix that.

Speaker 7: um so we can.

Speaker 7: i mean,

Speaker 7: we can tell the agents ah no go through my cold base,

Speaker 7: find out what the next thing thing to do are creating tasks for me at clicicbutton.

Speaker 7: i accept the quest under some some agent,

Speaker 7: the desert for me,

Speaker 7: which is fun.

Speaker 7: uh it's cool,

Speaker 7: but the problem problem here is that okay,

Speaker 7: now i have twenty stuff going at the same time,

Speaker 7: and i have two babies sit each and everyone,

Speaker 7: and that is completely exhausting.

Speaker 7: so you could ask,

Speaker 7: okay,

Speaker 7: we have agents.

Speaker 7: agents are amazing as we've established.

Speaker 7: so i can just go ahead and tell the agents to do it for me.

Speaker 7: now.

Speaker 7: it's called software factories by the times were just doing stuff autonomously,

Speaker 7: so you can give it like a general sense of what you want.

Speaker 7: it would sporn d and orchestrator and break it down for you and run it all in well,

Speaker 7: in this case,

Speaker 7: it's a local um container.

Speaker 7: so if it is just dyand isolation,

Speaker 7: and i don't need to do any babysitting,

Speaker 7: the asian does it for me.

Speaker 7: 对，嗯， on top of that, you have loops, which everyone really likes right now, so you can do cool things like just tell that to, i know sccan twitter for cools, if you wanna build or sccan to get get up, and it will just create this nice features for you in the background, without tunion to do too much up front.

Speaker 7: 对对对， the bad thing is that i mean, this is nice, but what do i do when i have again? I know i would take twenty, but even five things to review at the same time, it's very, very hard to get that out of the way.

Speaker 7: 嗯。 so what i did is create um this, like review view it. so i get the entire changes i need. and i get what file changed, if you want to go like file ile file, or can get just get visual evidence. so you can actually see videos of what change or photos of what change. so you can really quickly understand what you want or even run a few um instances in parallel and just speak the best implementation out of these.

Speaker 7: so now as we have this, we have um visibility and we have autonomy. it's prety cool, like we reised like a pretty good place in raising the scening of what we can do evagents, but it's still pretty exhausting.

Speaker 7: i mean, it was fun, but now i have to ask myself, do i really need to be the only one doing this? it's pretty lonely to just talk to agents all day and trying. and you know, if you look at um what a team does, for example, if it's just me, that's fine, but you have a team. i might not even be the right person to review what's going on. so i need something else, and that's something is collaboration.

Speaker 7: we can just have other people join us. so the way that looks in agent craft is that i can create these um h called no war halls, warwhom, whatever uh host them. it's like complete his local, but you can do like tunnels and people can join you. and so i can create a room, and i say have a practy designer in my team. in this case, it's my wife, and i can i prove her out a wife fugi met. so i can i prove her and she does her thing, she designs whatever. and i as an engineer can not only see that that's that ter. i can not only see that on my map, i can actually take that and follow off off of it, prompt yeh that just like her. um so i can take that design, follow up on it, implement it while she keeps working.

Speaker 7: and the cool thing here that not not can can hhand off for four can kind of have this shared work space ace, not not really bound anymore to get i kind use gget, but i'm not to rebound something. everyone is working together. you can sider. there's also some kind of checks between not only between the people within the room, but also between agents. so you can have really nice notice board of exactly what each person is doing, what each achating are doing are all aware of each other, and they can collaborate much more closely than you normally would be able to do.

Speaker 7: so for example, here i can say, okay, i took off this design task. so now now have a pop, but way this is also from devices. you can have like mobile and and i fleg telegramc and new telegram, whatever float your boat. so we have visibility and we have autonomy, and we have collaboration. so we managed to raise the feiling. it's now possible, especially if you're power users to do radicall stuff and use more agents and do more things.

Speaker 7: but the interesting thing here was, and if loolook at the feedback i got, it's not just that power users were able to do more. the coolthing for me was that people that had even nothing to nothing to do vagents or nothing to do with productivity. uh come and say, okay, my kids really love it and do it all the time to to work as strate agents, which it didn't do until now, uh or someone that fununt out of colleft, because he plaay starar craft. and now it can certainly do be productive stuff with it and a much of other examples of kind of bring communities that so far were kind of behind into this new world of productivity.

Speaker 7: so we need to raise the ceiling, but we also need to lower the floor. we need to have some kind of way to take, no, probably it around ninety percent of the people in the world and bring them into this igenentic future without having them burn out.

Speaker 7: so again, i took it too far. um and as this uh curtly named TVD, um potentially loopers, the idea here is, okay, why why do we need something like? you know what craft, which is potentially very complicated, especially might be even daunting if you're not into the grand ulity ity of what father are are, or what complex the UI is and kind of try to find a lower ddenominator of what makes stuff accessible, especially now that agents are much more capable.

Speaker 7: so instead of in onarity or or ving, an exactly what files theyy're woron on, maybe i can have it more of a mobile game level, have projects that they keep returning to and use looks. and our stuff as that rely on agents being more autonomouous to have like a very basic level of i prompt. i get the result. i don't really can't perhaps about the way i can't see it if i want you and have this really nice, interactive and full experience that makes me want to come back to this.

Speaker 7: 嗯， and obviously, and this is one visualization. it can can really resonate with some people. but might not presenent tive others because we're not all the same and what may appear like amazing and easy to one person can appear very difficult to another. uh but uah, i'm trying to look at the way to make this more customizable, try and understand what it is that you, as a person need in order to understand. and want to use these tools. so this is a very, very early draft. um but if you want to use it, please let me know.

Speaker 7: so we're the boallonenic. i think ''ve establish that pretty volunteer now. but the skills that we need to stop being the balloonic and enabled this new future are only with us. so we are the bonneck, but we don't have to be, and i hope that your can take what's relevant to you in your life and apply it forward.

Speaker 7: so i would ask you for two things. one agient craft is already out there like you can just index install. ageent craft is a website. um do your thing thing, this new thing that i just showed you popotitially uppers is still experimental. i'm so trying to understand the interest in what people are actually looking for to get the floor that they want toward the seeing that they want. i'm still try here and express your interest and tell me what you're looking forwso. i can know what to do do this.

Speaker 2: and thank you.

Speaker 9: see you in good fend of stulied search. and today, my talk later use a signals die at retrieval boundary. so look into uh what are agents? essentially, my agent feels what is the cause of fiels in retrieval, particularly and how to make actually signals cross the trivial boundary and how to make your agent basically unconfair.

Speaker 9: so uh, pluit started. but what is an agent and agent does an agalym that has agency to reason involve those uh interact with digal world retrieve uh the memory to complete the task. but major loophihere is missing is a learning. it should also learn from what to work and what it didn't work suppose.

Speaker 9: um if i have to uh explain what is asient and i can explain through occasient. so if i have to explain uh, what agent does, i had explained with react agent. so basically uit from the agent ent, uh, execute it in a look, uh, calling to a reree al research and then pause when the dusk 's company. this is very basic react architecture. one thing that is missing is how to make agent the noledge on the outcome. so agent keeps feiling at the same task.

Speaker 9: k. gotland report said eighty five percent of air takes failing chin, concerts and mkancies to in a deflight repreport. the problem came out to be most of the time is that tratree was static, seventy three percent of like my fierce because of the hieve, not generation and contect stuff. so i have n't post from around sure. i sht the exit to your paincorn said, uh, we have been optimising for the wrong thing you are paying and not for for are opwrwrong things. we have been optimising for the wrong things we made wrong answers up here faster and cheabird had before got to make greicree one loone.

Speaker 9: so why this this does does doesn't matter again. the type coomnize agents are not outcoming from. so there's n't missing milar, twtween, valils and action. your obsoability has all the traces. uh, all these staff that catch a absoilities. the star that catch every dounle, every airland, accomlition, every exceptions, you eva swijujuges ges, whether ffind that output was correct wrong, basically pass or fear, but these evalls are not reflected in agent context skills empdifies by agent action in any ways.

Speaker 3: 那个一一一个月，十一个一个个的一的一个。

Speaker 9: so the asian doesn't has any access to why yesterday is runs. h runs postor fear eva signal dies in the dash ford isn't n't missinginga.

Speaker 9: a system that consume traces absorb iraad and can brt both into the tree without its for future runs.

Speaker 9: so there's a manual improvement, tax and engineer actually has to sit and see if the email and also will be perform well, realiate the proms redeploit.

Speaker 9: a great you. they agreed to expense a model restructure, took our nets ts or ying to the custom models. why? i current memories feeling why memories was designed to actually add respese. but um it's not so they see uh what we we having as a current system.

Speaker 9: uh and current memory is that they basically store user differences, profile conversation history or long ononlived personalization, such child experience ces no acself improving the existence for production.

Speaker 9: if you see uh already existing approach in the uh market, market is uh launching um does mannal, which does emptifed references ua virtual signal is as embedding simility.

Speaker 9: 我们有没有 does it learn for mark could not. um so we have come up with something a new day discore, which is a similarity bit vadted by how useful it is for the agent to execute the task.

Speaker 9: it has actually the history of asked prieces and past items. so we came up with agent archies that that agents with trenn that experience. it's a rentern yare that that much agent improve from experices without free training, fine tuning or manual prom training.

Speaker 9: how it's a little different from compilying like the s pie because you big in all the lessons in the prompt here is actually improving because it's it is executed because so but again, introduce uh utility score. so you do not retrieve why he would retrieve resimantic and laded to the cutting dusk waited by whether those memories have historically helped or heg.

Speaker 9: the evi outcome becomes a first class signal in the dustory will reranking and not just sport cainful. one of the heat things is it. she ate memory as reason, not as facts, static fact with no conteicts and no history, but reasoning like suppose. uh if uh if theif, if is uh customer er support ers, uh looking for refriend, it will not only say y user er fers ers h uact user prefers uh it to be called by a short name. it's actually a reason about quite like like if somesomeone ask for refund, you should check the settlement before refunding it, so that the the the customer doesn't get be refund twice.

Speaker 9: so read based on usefulness, ness quantts is upgraded based on dusk. so this is a very big thing um because most of the agents here with context stuffing, and this has been rourought up in the ast ast, learn from history and reason ing. and this has from benchmarks.

Speaker 9: so uh, we have benchmark the our memory system reflect with on tol bench, which as actually measure if agents have followed a policy ban are com't. so we have seuuh uh uh, the performance improve from sixty six to seventy six percent without making in skills and wet skills. refleperforms ces eigheighty. uh, once it are uh enough memory like ten uh memories.

Speaker 9: uh we do is be beacon, the uh reasoning and the understanding into skills so that your agent always remains updated. what happens most of the time? yeah, seen suppose you have a product s cul agent. so there's a system. there's a column system proong even though that column is no useful anymore. it remains of the system problem. so there's no system right now that can update that. hey, there's no column uh right now call this. so really probably you shouldn't entertained it in the future. and this is possibly with skills that because ages always uuses calls that skill, updated skills all the time, and the similar behavior has mismissey and jupty.

Speaker 9: five five point for we have also benchmark it on agent dictators with essentially test a modest ability to reason plan unuse. those also extended multistep twks flows rather than measing ing stastatic QNA. so you can see here uh with the suppose the human last exam, uh with right, you get for desame point uh, the suarpoint uh with mememory stestem because h, with refilmememorsystem uh, this to sixty one, five, three, three perent. so this this kind kind of rend d shoore in another uh aggentic bbench marks um as well, like uh big cold bench, like long TV eccrow.

Speaker 8: so okay, hello, okay, before i get started, you guys, this is a huge key net room. can everyone like comford because i'm talking to like eempty roand disperse y, we'll do me a favor. i'm setting thirty minutes on you. all of our secrats. i can see you still. thank you. thank thank thank you. thank thank we're 're, just na chat. it's a giant room, and there's five hundred of us this term was way larger than that. thank you. honestly, i knew guys had it in you. it's really not so hard. thank you. i also sit the back. i also worked during talks. i get it. it's totally get it. i did all day, but not for me. okay, i'm going to start, but i'm going to still point at you if you're in the back like you. okay, i'm sarah. i lead our engineering teams for AI at notion.

Speaker 8: um looking to my talk, it's about token town. how do you go from not go from AI pilled to AI poor? okay, um i know that today's all software factories, we're going to talk about that, but we're going to talk about how to do it sustainable play. 嗯, this is me. this is on my first day and notion in a very sweatdy subway. um like i said, i lead our AI teams that notion um and i negotiate AI contracts for a living my team jokes. i act like ana winter. so this is a nice, a nice image of me with AI and a winter here after a press article referred to me that externally, um and that's kind of ideidea. um um how you think about negotiating bebetwen en different vendors, um making sure that mamaintain taste for your company. i don't do it alone. this is launched it one of our recent launches. this is just a subset. any good engineering manager points out that we have a whole company of people building this, that just the one that gets to come talk to about it.

Speaker 8: so we've been building a lot. um this is an example of our ai usage just just twenty twenty twenty six, and we've been really proud of how we've been able to grow that usage. and i'm going to talk to you about how you can build an AI native product and an ai native company. um but this is just to give me some credit that that we're doing, it kind of well 看. so for those of you that don't know, notions always been that durable system of record. it's always been the place where you can collaborate with your peers um. but today, that point of colcollaboration is a little bit different. it's not just humans. notions always been the place collaboration. and today, that collaboration happens between humans and agents, humans and humans, agents and agents.

Speaker 8: and we like to think about ai transformations going through this journey. and i'm sure some of you are looking at the slide in wondering where you are ai is a thought partner is when we all started tinkering. we all started just going to the very first version of chat y ptitin thanksgiving. when it came out three years ago, four years ago, and we started saying, like, how can i send this emails to my landlord to say that i ouldn't pay for repainting, right? then we copy pace, it entered entory email. eventually, we started getting to a place where we could use AI like like an assiant ant II able able mamay be execute individual task. that's how notion AI really took off in the beginning, and it was able to save employee time. but it functionally was limited in its capabilities based what humans acid to do. ai is timates is what we were really excited to launch almost a year ago now, but this is two in many products. we can do repetitive work and think about a process and have ai dudo that process. what i think is really interesting is when ai actually becomes that critical workflow, where processes are interfacing with each other in new of entire systems running.

Speaker 8: how many you guys feel like you have? ai is system down? are she said you came up now coming? great. none of you exactly. we have found that no one has figgered out how to do this. well, eighty eight percent of people can't even get past ais in a system. and why is that we've at these? is that notion, it's because there's too much cylude data and not a sturable system of record for that point of collaboration. and we believe that for your software factory to work for your company to work and for your systems to work, you need that durble system of record. and that is nocean's mission.

Speaker 8: so doing that is expensive. you see a lot of companies that try and commit themselves to this vision. and these are just a serious heheadinines with within a week of how that's painful. so you can put all of your money into a process to try and make a system. and you end it feeling like this, right? you end up using a blow torch to light what is actually at large cigar. but you kind of get the idea. cost is a structural barrier to entry. it makes it hard for you to serve products.

Speaker 8: it makes it hard for you to build factories.

Speaker 8: and it is ultimately, i would posit one of the largest reasons why things do not happen at scales successfully day.

Speaker 8: and i would argue for anyone working in an applied AI company at something for them to be really familiar with to understand the trade offs that they are making to build durable and exciting and enlightening products for their customers.

Speaker 8: but that's that really how the market is today, right?

Speaker 8: i'm not going to name names, but you guys have search engines.

Speaker 8: she can figured out exhibit a reasoning model, gets upgraded. amazing that poor token pricing is the same. what's not to love you. try it out, uses three times as many Alpaca tokens right exhibit be a model. it's upgraded, but it has an entire new digit, right? whatever marketing system that model family likes its brand new it's forty percent more than its predecessor, which is being deprecated in the next four months.

Speaker 8: these are real scenarios that we face. all of you are nodding because these are common pretty much monthly now, but here's the problem. are you growing forty percent in that time period? are you making thirty three percent more revenue? no. so how do you navigate this system if you just ought to upgrade your model in everything that you're doing? you're giving one bad deal, either your customers or your investors, depending on how you charge and where you get your money.

Speaker 1: neither or good fortune.

Speaker 8: five million companies have the capability to navigate this. they can hire large consulting teams have durable teams on their own and build expertise and handle navigate these tradeoffs. most people don't everyone else has no ability to negotiate with leverage, and they're stuck in these scenarios right? part of my job as that an Anthropic joke is to think about advocating for the fortune five million, the non fortune five hundred companies that don't have the mass to have leverage and negotiate, but need to think about how, and i'm going to share some the lessons that i've learned when i have kind of large amounts of traffic behind me that i think scale to those who don't.

Speaker 8: this is probably less of a secret. now, then it was when i started giving talks like this, maybe four months ago. your supplier is your competitor. i know very few people who have convinced me that that's not true. you always be getting a bad deal, especially with someone who builds them natively, right? sometimes the cost of good service is extremely different. you're basically they're serving a first party product and then you're buying those tokens at a huge surcharge, and then then selling them again another another surcharge. that's not really value. you can defend you're getting a really bad deal. and if you tie yourself to one provider, you have no exit. if you build an AI product that you're selling with this structure, you are crossing your fingers and hoping that you are a viable business and do not encourage that.

Speaker 8: this is really interesting Dylan and some analysis posted this. i think it's says eight hours ago, it wasn't at this slide probably a month ago, they purchased this subscription plan, and they just highlighted right how different what frontier labs charge customers for first party products are versus what they sell. it's a bad deal. don't play this game or try and let me know how you win. i don't recommend think about everyone else think about what that structure means and where you have expertise. i don't think that that's winning on the token economics. i think it's about product. it's about building data fly wheels and understanding your customers better than anyone else understanding when you need capability when you need low price when you need latency improvements, i promise you you don't always need. it's usually the lowest, but the most capable model out there and then build compelling, UI and orchestration, and also use some examples of that to justify the cost on the bad deal. tokens that you do resell the job is not to train. i mean, some of you might be training the best model, and i'd love to serve it and come talk to me afterwards that most of you are not doing that. stop trying to win that game and think about the best product that uses many models, help your customers help your team bet on the frontier, not on the lab, and we'll talk about what it looks like to do that.

Speaker 8: this cost for capability bridge second tradeoff is actually really intense. cited it. LLM came out with this memo a while ago, maybe two weeks ago, i loved it. the idea that for the economy at large, simpler models might be the most cost effective productivity augmenting pathway. they talk about this bifurcation on frontier versus every day usage. i really believe that. and for every product, the definition of frontier versus every day, the definition, saturated capabilities or model capability, overhangs depends on your expertise on your product. no one can replace that in not all traffic is equal. it is a huge miss to send all of these to the latest OpenAI model. some of these absolutely large skilled data analysis when you do it on Notion, we'll recommend OpenAI right when you treat an email in box. if we're charging you to do that on OpenAI, we're ripping you off in ourselves, think about where your traffic patterns are. and then think about how frontier lab model providers are structured today. i mean, it's functionally an oligopoly, right? and that's fine because they're racing to the top. and i think the top is really hard and really important. this is not to say that products don't have a place for frontier difficult task. i want everyone to not misunderstand. that's not what this talk is about understand when you need those task, and it's not everything. the problem with those task are is keep in mind how pricing is incentivized. you can figure out who these players are either. you are the best model, everything above what AI can do today is your market. you can basically price it as high as you kind of want. if you're slightly behind that best model, all you need to be is like a dollar per million tokens cheaper. and you have the rest of the market. you know that economic theory about gas stations were the best gas stations are the ones that are right next to other because they cover east and west. the most yeah, it's the same with model pricing, which means that price does not correlate with capability growth. so for this complex task, understand what capabilities you need, but be the expert in what complexity is and keep in mind that who handles complexity changes.

Speaker 8: often times you'll see applied AI companies really be outspoken on marketing with a specific lab. that's always kind of a red flag for me when they're not model agnostic. because if you look at the graph, it basically shows that they're behind every month, right. the new model in the new model provider of the best frontier capabilities changed. and if you hit your ride with one particular provider in exchange for reference, a larger discount, i'm you're doing it to service your customers like half of the time.

Speaker 0: right. so really think about if that discount is worth not actually having a frontier product. and remember that, that optionality is your leverage, if you don't have the capability to walk at any point, you are stuck. and again, i think that's probably the most expensive decision. you'll make regardless of what discount you get or the engineering work to have model interoperability one option to navigate this, a state model agnostic have different models and capabilities in your system. so that at any point, if pricing seems unfair or untenable, you're not out of business notion's auto model. does this really well? we have stated the our models available always, but we also have an auto model there at the top that handles about seventy five percent of her traffic, right? we have the ability to switch between models in our product, and we also offer it to our customers that they have access to these models without vendor locking. that's part of our AI Switzerland approach. you guys left taking photos of slides. this is the slide. okay. model logistics playbook, this is how you do it build from multi model. it is hard to kill the guesswork, which models may transcribe. i understand that we invest in that technology. it doesn't even have to be per thread. just think about your harness's model interoperability. think about the cost per capability per second, not just the tokens, here's a great example. we posted this review when we announced our partnership with parallel as our web search provider. if you were to look at just length of single call, all or just cost parallel might not be the cheapest. but if you have expertise in entire web search trajectories, you'll see how it differs the granularity of this level is what lets us make the best decisions for our customers because we understand all of the tradeoffs on entire trajectories, not just single call. switch fast. and often, i think we talks about that and give them something back. that expertise on use cases is also very valuable to frontier labs.

Speaker 0: we find that our evals in our early access program partnerships actually help us a lot with frontier labs and is something that we can exchange instead of extraordinary large comments.

Speaker 0: and i don't think the discount is ever worth the lost and opctionality.

Speaker 0: that's a perspective.

Speaker 0: you can choose to keep her,

Speaker 0: not the second option is moderate.

Speaker 0: task,

Speaker 0: understanding open weights place there.

Speaker 0: open weight models are really strong enough to handle these tasks in the possibility to oral on on top of them has also kind of expanded the upmarket growth that they can cover.

Speaker 0: i view open weight models is basically lowering the barrier to entry on cost for our customers.

Speaker 0: and they also give you negotiation leverage.

Speaker 0: so it's kind of a credible alternative that's putting that downward pressure on pricing that if there's an allogopy of tour three providers at the top is unavailable right now.

Speaker 0: otherwise, i think came me two six was probably the first time that we really saw a model that i performed five, two GPT, five, two, two LM, and five two now was another five, two bomb shell in the villa.

Speaker 0: that also probably just best here, but it's no longer the case where open way models are good for just sf t on small tx really think about without our rel if they're capable enough for what you need.

Speaker 0: and again, don't just think about externnal benchmarks be able to have expertise on your system.

Speaker 0: what are your tool errors, what the actual latency that you need?

Speaker 0: right? here's an example of a benchmark that we pushted a little bit steill on purpose, right?

Speaker 0: but you get the idea fill if a based tenfof this showed this slide once, and i've stolen it ever since.

Speaker 0: thank you.

Speaker 0: are you here? buddy?

Speaker 0: okay,

Speaker 0: oh, chat.

Speaker 0: hi,

Speaker 0: hey.

Speaker 0: well, he could come up and say it better, but the idea is that you don't have to be at the top, right?

Speaker 0: i'm not trying to make a case that open way it is the best model out there.

Speaker 0: the case being made, however, is that the gap gets covered eventually.

Speaker 0: so if the task that you're having today are good enough than in six months, they're probably covered by open way.

Speaker 0: so be prepared now.

Speaker 0: and the last thing as cps of our gps, we've ve recenlalaunched something at notion called workers.

Speaker 0: i don't think that the gp was necessary for every job, a lot of the jobs that we have are actually serving um discrete pieces of code like you don't need an alymp to turn a csv into a pd f.

Speaker 0: you don't need an alymn to talch a notion to a calls.

Speaker 0: if we have ACLI, you definitely don't need alalympt to determine stic eal queries.

Speaker 0: this is where people become token poor, very quick.

Speaker 1: 没没有没有。

Speaker 0: and i think the last option here besides open weight c in and opoptionality is actually governance.

Speaker 0: there's a lot of egi governance. um one is visibility underststanding, who using the the atata underananding maintainability and control.

Speaker 0: when you have model optionality, you can offer a lot more to your customers.

Speaker 0: i'm uan example of how that governance works a notion.

Speaker 0: so final tips, again, think about architecture, think about open weight and build value that transtends tokens.

Speaker 0: so we're going to depart token town. i know i said of welcome to token town. and when going spspend next next minutes, really thinking about what to do next.

Speaker 0: so i think the challenge of the next six months doesn't have to do with capabilities.

Speaker 0: i think it has to do a security.

Speaker 0: let's start there.

Speaker 0: there's this concept called the lethal trifacta simon wilson son. i think, crafted this,

Speaker 0: if you have access to private data exposure to interested content, whether it be through ingjestion mcp, email and the ability to x to communicate externally, and that can include like payloads in a web search.

Speaker 0: the second you have that system yyorks wissing risk.

Speaker 0: and in fact, the more autonomious system is the more unsupervised this risk is,

Speaker 0: i think that this is what builds valuable product, not just capability,

Speaker 0: same with sandboxes and computers.

Speaker 0: we talks about this, but it really is something that builds better determinism in your product and also better token economics for your customers in multi agent, orchestration understanding what agency do and what persists.

Speaker 0: i think perperstence of enterprise knowledge is something that's actually really not discussed enough starting to be with some recent launches.

Speaker 0: you know, there is audio, so don't have your workclothes look like this.

Speaker 0: and i think this is where most suoftware factories are today.

Speaker 0: right? it's like actually your entire engineering time just spend time babysitting the factory, right?

Speaker 0: i mean, i get it are started off like this agent orchestration is one of the most difficult tasks of making factories work.

Speaker 0: so okay, this is me telling tea pain to tell people to buy nation AI.

Speaker 0: and the reason i included this slide is i am going to sell notion for a second.

Speaker 0: it's my job always be closing, always be selling, always be hiring confined me.

Speaker 0: but i'm going to talk for a second about how nation does this today.

Speaker 0: we already have the ability to inspect task.

Speaker 0: and you can imagine any task that you look at um in a notion document, you can have claud actually go ahead and scope out what you need.

Speaker 0: we've launched this managaging capability today.

Speaker 0: so if i go ahead to the top of this task, i can actually ask claude agent just go about the task, right?

Speaker 0: ideally. it's working in your seettle actually populate an entire spect of what needs to be done in this example. it's not ready.

Speaker 0: it's going to ask me a question.

Speaker 0: keep in mind, this isan a mark done file.

Speaker 0: this is an active document.

Speaker 0: let's say, i don't actually know the question, and i go ahead, and i ask my team um what to do,

Speaker 0: imagine that you can kind of tag in your team into these systems, m jasor pm.

Speaker 0: so in this example, she doesn't know usually she does, but motly agent organstration is important.

Speaker 0: maybe cloud code isn't the best accost for our voice.

Speaker 0: but decaagon is right.

Speaker 0: you can ask decagon agents or proud partners to them as well, to collect three data that you need.

Speaker 0: okay.

Speaker 0: in this example, we think we know enough we're going to go ahead and actually iterate through some of this fflow.

Speaker 0: and a skip had a little bit.

Speaker 0: we asked her till what we needed.

Speaker 0: he replied,

Speaker 0: again, it's a collaborate file and not just to mark down,

Speaker 0: and we can have cloud actually go ahead and spent at the pair.

Speaker 0: hopefully, this is looking a little familiar.

Speaker 0: now this is kind of division of software factories.

Speaker 0: it's what we're trying to host.

Speaker 0: okay, claud put up a pr.

Speaker 0: maybe that's not enough.

Speaker 1: maybe i want to go ahead and ask cotex what it thinks great found too.

Speaker 0: issuous, you can think about this scaling in an actual factory.

Speaker 0: so today, in notion, you're actually able to orchestrate these agents together and you're not committing to a lab.

Speaker 0: you're committing to the concept that ai is augmenting and automating what you do.

Speaker 0: this is real.

Speaker 0: i asked gief if i could post this, this is how it works today.

Speaker 0: internal in ocean, almost all of our polish and large feedback like this is actually coordinum through our software factories, both in terms of writing to the right teams and also having cutting agents take the first step verr cell.

Speaker 0: does this as well from staging to shipping to closing, and we see massive araoi gains from our customers.

Speaker 0: it's over three minutes saves on a given task.

Speaker 0: imagine that at scale.

Speaker 0: so i think we're trying our hardest to think about the factory length.

Speaker 0: we cannot do this without opportality,

Speaker 0: and we cannot do this without conviction that we understand what models are required for, which tests.

Speaker 0: it's really wild out there.

Speaker 0: you guys i get it.

Speaker 0: the market is really young.

Speaker 0: it's exceptionally oppic.

Speaker 0: it's moving fast.

Speaker 0: i'm super grawful for committeties like AI engineerto, bring totogether and talk openly ly about these things and how we navigted.

Speaker 0: um i think we owe it to all of our customers to get it right and to be critical thinkers, how we naviviate this together.

Speaker 0: i'm chchonicically online.

Speaker 0: fortunately, um you can always DM me on twitter.

Speaker 0: you can email me.

Speaker 0: you can find me after this.

Speaker 0: you,

Speaker 3: you thank you for yepping with me and thinking about this problem and have a good deck is for ah just just got so much in there.

Speaker 4: it really helps write better code and structure your code better and helps with architecture spending up new services for and for us on the agents team really helping this design and build the the corage ge and opp.

Speaker 4: so you'll see throughout this presentation, sprinkle, then um how afffect on our team uh has paid of individent.

Speaker 4: so we really love a effect you at open gove and h.

Speaker 4: we encouraother other folks to try it out.

Speaker 4: and yeah, let's keep going the effect native loop.

Speaker 4: so originally, we are on langrarah.

Speaker 4: and that was fine until the team really started to scale.

Speaker 4: and our use cases started to evolve.

Speaker 4: so we decided to move over to our own kind of affecnaated agent loop to have full agency over this agent loop such that if we have complex use cases or features that we need to build,

Speaker 4: we could kind of get in. we had full control of the agent loop, but not only that, but now we're fully on effects. so all the cool things you get, what effect is now propagated throughout the entire agent loop, like the tracing structure can currency to logging. everything is more fine grain control. and so it really allows us to really unlock the full potential having our own agent loop from the ground.

Speaker 4: so another thing i wanted to mention is you'll see a code example. this is really the basics of the effect loop that we're using. we're using this thing called the effect ai package. and in that package, there's this thing called a chat and a language model. so for the chat, you can instantiate a chat for example. and then you could stream text using that stream text function, you could pass in a prompt and what's cool is with a language model under the hood we're kind of doing dependency injection. we could pass in a different language model if we were to hot walk to another one, for example. so really just having full control of our own agent loop, just kind of give us all the levers. and it really just unlocks the full capabilities of the model and for the team as well to have full agency over this loop.

Speaker 4: the thing i wanted to mention is the agent to agent protocol. so here on the agency, we've had a lot of success with this protocol. so this protocol being the protocol that google created kind of an open protocol for agency in our communicate. but we found this very useful for defining our agent routes, for example, in the back end and our model and our schema to follow this kind of agent protocol. so we model. so for example, there's this thing called an agent card. would you see here? and it's got the name of the agent, a description et cetera, right? and having this kind of a rigorous protocol. this rigorous spec really helped drive our development and drive alignment. because you know, all we had to do was align with this spec and follow this spec, and we knew that this was kind of the contract that our front end and back end both consume and produce. so this, i would say also has been very helpful for us. and what's really cool is a2a has a lot of extensions, right? so you can extend the protocol with like metadata. there's also a2ui. so lots of fun stuff with a2a protocol. this is kind of what's worked for us. so sharing that with you folks.

Speaker 4: feedback and evals. so here the quote is shipping is the start not the finish. so what we do here on the agenscene is we have kind of multiple ways we do evals and collect feedback. obviously, you know, will have folks call in or email us or just let us know and tell us, but the main way is we have the thumbs up and thumbs down mechanism. and here someone is able to tell us, hey this work really well. this was a great response or that wasn't a great response. and that signal we take and are able to iterate on and we can take it back and help improve the response in the future. we also have automated evals. so in our CI, we have evals that run against real completion. so we could test the prompt against. hey, did it hit some tools? did it do what it's supposed to do. and that also helps with our accuracy. so those automated evals in conjunction with collecting feedback really help us improve our tools, our skills, our harness, and that's really how we are able to iterate so fast and so quickly.

Speaker 4: humans in the loop. this is a really cool feature. we built where we deterministically in our agent loop. if there is a tool call approval required, so if an agent tries to make a tool call that it needs human approval for it'll show this UI and the human can click accept or reject so explicitly reject or explicitly accept the action that the agent is trying to make. this ensures that you know, we're building trust and also ensuring that we're being safe, especially when the agent trying to do a mutating operation and always always always making sure that humans are in the driver's seat. sandboxing. so another thing that we worked on, going back to the safety slide we just saw, was whenever the agent tries to execute code or tries to create files. it does so in a sandbox. so we gave our agent sandboxes such that it could spin up these sandboxes on demand and it could use those sandboxes honestly, right? code execute code create files. and it's kind of this safe temporary isolated space such that the agent can take action in there and we don't have to worry about any risk to our production system. it's really, really cool. they also get torn down at the end. so in this example, i said, hey, create a PDF for the folks of the AI engineer conference on twenty twenty six. and allow me to download it so i could share with them.

Speaker 5: and you can see that the agent created this fighting slop would slop.

Speaker 6: my name is byband. i'm want to talk about something that is a little, i would say maybe a little silly at first. i'm want to show your team's engineering practices really quickly. we do no code reviews. we require every engineer to work on things in parallel, and we have no standardization how people do ai. and i know immediately what almost all you are thinking. we're probably as sumor wise a start up, and i can guarantee you i'm clearly millennial. so what do we actually do at our company without code reviews? well, we built about three years ago, we decided to build a programming language. that's something that has absolutely no room for slop. it's something that has to parse every single time exactly the same, same way, something you can't just change a year later, or a month later, because you made a bad design decision, you have to be correct for the last three years. we've been an unsung war against slop. and when i first met this enemy, i went to my great mentor, slop sue, and he taught me something to defeat the slop, we must become the slop. so we began and we prepared and then we started winning. so when we think about it, what is slop? slop is just any code you don't read and whether you admit it or not. this is the least amount of slop that your code base will ever have cherish it. so we started fighting back against the slop when we started fighting back with slop. so how do we go ship a stable programming language with these engineering practices? while the first skirmish we ever had was the skirmish of standards. the hard part about hiring great engineers is, you sadly can't tell them what to do. some of them want to use claude. some of them want to use cursor. some of them want to use the latest thing that they just found on hacker news. so instead of trying to hold standards in our code base, we did something that is an invariant. we built an architecture out nd file instead of using claude md. just picks something that every model can just understand. this file has been incredibly small, and it can only have things that will not change for months or for years. in our case, it's the layers of the compiler you go deeper into the compiler, tell the devs that just talk to at least one other person that slows it down a little bit. so now we had standards. so anyone can use whatever they want, but the real foe we face was actually the battle of design. everyone here knows that you have the right perfect design docs. and we have a very simple rule in our team. code can be slop, writing cannot. and of course, if i tell every engineer this, they write beautiful writing, and they hand write everything they don't use AI. well, sadly, not. so we built the design doc tool. this design doc tool is a replacement for both notion and github effectively for design docs. it has versioning, commenting, all the stuff you want. and obviously, we do this, people use it? well, sadly not. we built another tool on top of that. and this tool was a slack integration for that tool. every time a design doc was updated this channel got notifications, and what ended up happening is this channel became the most popular channel on our company really fast at two AM. someone posts a design doc, three people started reading it right away because it's just interesting. the most interesting stuff is design docs. not calling to change, but this wasn't enough. all of this is actually backed by markdown files and simple ci scripts. i make it treat like github without being github itself. so now agents can go do this. but the real problem with all this is i built this, and i hit a little bit of psychosis, and i started shipping ten design docs a day. and soon the team was fighting my slop. so we had to go back at last rule. this last rule was if you're going to ship a design doc, you require people to actually go read it. and with this last standard,

Speaker 6: we suddenly had design dogs that are incredibly high quality.

Speaker 6: but what about the battle of architecture?

Speaker 6: how do you have your code base converge?

Speaker 6: we build to another tool this tool?

Speaker 6: basically visualizor dependency graph internally with some external depencies as well in lows to watch the code base change.

Speaker 6: it has semantic boundaries individual packages,

Speaker 6: but it's more interesting is we can go build CAI tools that guarantee that certain invariants can't be broken.

Speaker 6: and what this does is when cloud build a new package or as a dependency,

Speaker 6: it's leaky.

Speaker 6: we now have CICD changing or a simple git commit history.

Speaker 6: it tells exexactly with break ak by by we're actually tually to american architecture ture change.

Speaker 6: we haven't ged our architecture and last three or four months.

Speaker 6: but as much as we might do,

Speaker 6: design dogs.

Speaker 6: and as much as we might have stable code,

Speaker 6: would you genuinely ship code without reading it?

Speaker 6: would you trust your team to go do that and think about a programming language?

Speaker 6: a progragramming language has so many invariance you you gengeners,

Speaker 6: you have closures,

Speaker 6: you have memory allocation.

Speaker 6: you have FA five boundaries.

Speaker 6: could you trust that system python as box twenty five years later?

Speaker 6: well,

Speaker 6: here's where we did something insightly different.

Speaker 6: what we did was we built a system that actually has agents constantly running and creating demo programs.

Speaker 6: we take these demo programs,

Speaker 6: one second,

Speaker 6: and we have agents try and spin something up from scratch.

Speaker 6: we then look at the entire cloud transcripts,

Speaker 6: see what tools you see what happened.

Speaker 6: and obviously,

Speaker 6: ly,

Speaker 6: as humans can inspect them.

Speaker 6: but more importantly,

Speaker 6: we can have agents go inspect them.

Speaker 6: and agents find what was was ls.

Speaker 6: what was bad and not just what was bad into intervworld is incorrecting language.

Speaker 6: but what's the three tool calls when it should have only taken one?

Speaker 6: and then we can go head and find issues,

Speaker 6: and we can have humans collaborate with these issues to figure out which oneare real,

Speaker 6: which one's are hlucinations,

Speaker 6: which ones are don't have taste as much as i hate to use that word.

Speaker 6: and then we can have agents go head and create fixes to these problems and go address them.

Speaker 6: and most importantly,

Speaker 6: instead of trying to just detect these issues,

Speaker 6: we can go once that further.

Speaker 6: what if you could find language features instead of guessing wall was good guessing.

Speaker 6: what school was good.

Speaker 6: you could go in AB taxi.

Speaker 6: you could forget out onees took less tool calls,

Speaker 6: which one took,

Speaker 6: which one made less errors,

Speaker 6: which one produce a correct outcome.

Speaker 6: and deterministically know what's going on.

Speaker 6: the point is you can start building data or and systems without ever writing a single line a code.

Speaker 6: and the thing that really i care about the most over here is not that any one of these tools is specifically what you should go build,

Speaker 6: but the fact of the matter is,

Speaker 6: in order to able the programming language,

Speaker 6: it won't have taken eight people.

Speaker 6: it won't have taken less than two years.

Speaker 6: it were have taken hundreds and thousands and tens of thousand man hours.

Speaker 6: and then you would still have a broken system.

Speaker 6: and today,

Speaker 6: we can to spend billions of tokens and make it work,

Speaker 6: and we can make it stable.

Speaker 6: and you two can go home and build these internal tools and these sloppy tools to make sure that your code basis can ship without really having to read necessarily every single line of code,

Speaker 6: because your engineers aren't going to.

Speaker 6: and i think we can start winning this battle le and sloloand as we win this battle.

Speaker 6: slop can be defeated.

Speaker 6: but sadly,

Speaker 6: i have a sad thing to say,

Speaker 6: i think we're still going to lose the war.

Speaker 6: and i think the reason that we're going to lose this war is because some of the foundational stuff that we try and go use itself is broken.

Speaker 6: how many of have use sidescrit？

Speaker 6: probably most of you hopefully at this point or at least your agents have so something run there？

Speaker 6: did you know that types？

Speaker 6: er's main design le is the strike balbalted，

Speaker 6: corcorrectness and productivity.

Speaker 6: and there's an astricts year because what they really mean is human productivity.

Speaker 6: and if you think about it，

Speaker 6: there are things you would never do in a programming language at the very core layer.

Speaker 6: if you were designing for a world where where humans never wrote a single encode and let me show you what that really means.

Speaker 6: i'm going to to write something and try and guess what this code does pretty safe，

Speaker 6: what about this one or even more？

Speaker 6: so this one，

Speaker 6: why do we turn things to strings when we sort them？

Speaker 6: this is just slop big into the language，

Speaker 6: whether you like it or not.

Speaker 6: what about this？

Speaker 6: i love this part of type script，

Speaker 6: and you know what my agent loves this part of type script.

Speaker 6: this is slop big into the language.

Speaker 6: and whether you like it or not，

Speaker 6: the systems will have slop.

Speaker 6: if you build using these these tools，

Speaker 6: i'm i'm sorry，

Speaker 6: wrong talk.

Speaker 6: but if you think about what javascript does javasscript exists，

Speaker 6: and then after javascript existed，

Speaker 6: we started building systems.

Speaker 6: so leare it on，

Speaker 6: we built coffee script，

Speaker 6: the entitship.

Speaker 6: and now we're trying build effect.

Speaker 6: but thing is the thing underneath is already broken and more so the way we right code is also different now.

Speaker 6: so why are we trying to pch something like this？

Speaker 6: why don't we just try and do something a little different？

Speaker 6: and i think what we might need if we try and go do that is basically going to be a made up language.

Speaker 6: so let me show you what bammo really can do.

Speaker 6: and when you start thinking from first principles，

Speaker 6: how you can try and combat swap from a very foundational leyer itself，

Speaker 6: i keep talking about not reading code.

Speaker 6: does it even matter？

Speaker 6: well，

Speaker 6: let me show you a new way to think about code.

Speaker 6: and this isn't to say，

Speaker 6: we all have to go do this right away.

Speaker 6: but what if every single time i looked at code loops？

Speaker 6: what if every single time i looked at code，

Speaker 6: what i really saw was not the code itself，

Speaker 6: but a quick little thing that could actually visualize all the code for me.

Speaker 6: as i clicked around，

Speaker 6: it took me to exact lines.

Speaker 6: the code i was link to.

Speaker 6: if i wanted to have a slightly brought of you，

Speaker 6: i could zoom in and click around and have it expand.

Speaker 6: and i could naviget my code basis more intereingly.

Speaker 6: i'm moof of this rurun really quickly.

Speaker 6: while while it runs shshoyou a different pipeline without any you ever reading the code，

Speaker 6: you know，

Speaker 6: i'm setting up south and haan agent loop because of semantic boundaries in there，

Speaker 6: i can expand this.

Speaker 6: i can keep expanding this.

Speaker 6: and because i hope that's too much slop with let that be slop and walk away.

Speaker 6: so instead of having to understand all the code，

Speaker 6: i can opt into what parts of the code i want to read and understand and go to the exact lines when i really care about them.

Speaker 6: if we go back to the previous pipeline，

Speaker 6: i was running，

Speaker 6: what if while it's running，

Speaker 6: i can actually get a full execution trace in a world where we don't read all the code.

Speaker 6: the only way to understand the code is actually by the execution trace.

Speaker 6: and actually，

Speaker 6: by seeing exactly how much time will spent on what parts of my program at any given time，

Speaker 6: if you want to go and actually track your program through，

Speaker 6: think about a slow，

Speaker 6: your program would be if you had to go trace everything in pipe on our typshriit.

Speaker 6: it's untennible.

Speaker 6: and the best part here is if you start from first principles，

Speaker 6: you can make this effectively zero performance cost 没有。

Speaker 6: not only can we make it great for humans，

Speaker 6: but because that's all belt for agents anyway，

Speaker 6: you can go ahead and make it.

Speaker 6: so that every single file has a tracing system that clod can navigate through.

Speaker 6: so laud can find what we're bogs，

Speaker 6: what were ears.

Speaker 6: and what we're ineficficenciciand and art opoptizing yhode without you having to do it yourself.

Speaker 6: and i think if we go sort thinking matter from this way，

Speaker 6: it's not so much about reading all the code，

Speaker 6: but it's more so about as a human understanding the system mature working with and the tools that you can build，

Speaker 6: can give you information about the system that ure working with.

Speaker 6: but i think there's another letter to it.

Speaker 6: we spent decades building.

Speaker 6: i tool link，

Speaker 6: and that think about how long it took before someone like me who does not know how to escape them to this day，

Speaker 6: can finally start using VS code.

Speaker 6: it was a beautiful day when that happened.

Speaker 6: i became a real programmer.

Speaker 6: well，

Speaker 6: according some people，

Speaker 6: i'm still not because i can't write vim code，

Speaker 6: but what does agent first tooling look like？

Speaker 6: i think we're all familiar with grap.

Speaker 6: so i'm not going to go and talk about it，

Speaker 6: but i will talk about a rip grap because ggrap should not be used anywhere if i want to grgrap through my code base and understand what it was，

Speaker 6: i would regrab say something like calculate.

Speaker 6: and i'd give me a bunch of code where everything was being used and maybe you be somewhat useful.

Speaker 6: but what if you could instead to start describing code and say it，

Speaker 6: can you described calculate for me，

Speaker 6: what if it came with all the doshinings？

Speaker 6: what if it came with the actual source code？

Speaker 6: and what if it also told you everywhere was actually use under the hood，

Speaker 6: we can make something uto be multiple tool，

Speaker 6: calls a single tool call of a sudden.

Speaker 6: what if the way you want to learn about libraries that you are using instead of have to do a web search？

Speaker 6: you just said，

Speaker 6: you could just ask for any external library as well and was just given to you.

Speaker 6: because when i first love start, ted learning had a code one valuable lesson. i had was the code is always a source of truth. don't read anything, but the code itself, the docs may lie the um the actual description or architecture file readme file will definitely like, but the code cannot lie. except if you're working on some we're recollectures. and then when you go down this road, you go from not reading the code to understand the architecture you go from not searching the code to understanding exactly what you're getting an everyone tool call. but what's the next thing you do? well, the last thing i do to truly understand code is i run the code.

Speaker 6: so what if every single thing you ran every single function you ever wrote was immediately available. you know, potiticco over here was immediately available as a simple CI command. so if i run add, add becomes a CI command that has arguments attached to it, and i can run it really quickly and see what happens. what if every single tool i had could be packed into a unit test, that's completely stand alone that i can just run without ever having to actually execute any of the code. and as alalc tools, a binary that has functions just bundled in some way. we can build really quick tooling where agents don't have to go through what's happening. everything is typesafe, everything is deterministic, and everything is actually guessable. and the best part is imagine you could build on any system, and your agents don't have to worry about deployments across windows mac and linux. and you can just target any layer you want. and it builds for any architecture, including wasm systems. all of a sudden as an engineer you're supercharged, you're no longer bound by what you can do in the systems underneath you are preventing. you can just move very fast. you can move at agent speed.

Speaker 6: but a lot of the stuff that i've been talking much to the state has been about tooling. what if we tried to fix some the real sense of javascript, some of the stuff that is deep in the language, not the sort stuff. but i mean, more important stuff like errors. have you seen error handling be beautiful ever other than rust? um what i see agents do over here is you do try catch, and then they keep nesting try catch after try catch after try catch and imagine they give up in some sort of lot, log some error happen and deal with it. what if we could do error handling from very first principles, what happens in that world? well, i showed you add multiply function. i didn't show you divide. divide is dangerous. it's spooky. so let's go look at divide. you can see over here, divide, throws a division by zero error. but what else happens? the function actually knows that it throws division by zero error without you having to write any any code, if i go up to the calculate function, which at some point calls divide, this function also knows it throws division by zero error. so error types now get inferred without you ever having to do any guess. that means that you catch or handle errors, we can do exhaustive guarantees, and the compiler can prove that you have handled the error or not handled the error. it's no more guessing. there's no one knows it's guaranteed to be proven. so if you want to ship an api where guarantees, it never throws. well, this system is broken because it doesn't meet the constraints. it has two errors that you're not throwing. if you want to go catch that. well, i can write the code for them a second, but you can start catching certain errors. i mean, i'm going to return a sentinel value for now. and now this parsing, which previously threw division by zero error is now guaranteed to no longer throw the division by zero. because if i catch any exceptions in here, i return a zero value, every single time, the compiler and the tooling can do a lot of work for us.

Speaker 6: and we're already used to this in our code bases. we many of us probably don't know how compilers work under the hood. and we trust them code is a matter of trust. the reason that we don't use code blindly is, we trust it, and the systems underneath them. we don't have enough traceability. one more thing. but before i tell you all to go write a bunch of bammo code because i've been there. and i can tell you what someone would tell me, they said, hey, using a new programming language is not going to solve all your problems it's just going to come with a whole slew of new problems. so we said, i think we'll lose the war on slop if we try to ask everyone to rewrite all the code in the world into this new system. so what does a solution like that look like where you don't have to rewrite all your code? well, what we started to do was we started to think about that about two years ago. and we said, what if you could use bammo, not just stand alone, like i show today, but from within any existing language of your choice, from python to typescript, to rust, to go to ruby, to java, to anything new that comes up even after it. what if every function in bammo is immediately accessible in the language of your choice. in this case, i'm calling the bammo calculate function directly from python and it's completely type safe. not only we get calculate, we get calculate in typescript. in case some of us want to write correct code. so bammo while it has no function coloring, it does give you the benefit of having to do whatever you want across your code. but what if you want a little bit cooler, what if you started passing around lambdas across language boundaries? i have a function called with timeout this function times out after a certain number of seconds. and if this work doesn't complete, it's guaranteed to no matter how long it takes. in that world, you can even pass python lambda across the bridge. you can pass generics across the bridge. you can pass closures, it should just work. so engineers don't have to go fast with it. and more importantly, so when the agent does something, the type system never lies, the type system becomes the absolute center of truth that prevents invariants from entering your code base.

Speaker 6: and what i really wanted to talk about today was not any one specific thing, but it's a general concept. you can build incredibly complex systems without traditional systems like code abstractions. you don't you can work in things in parallel and you can use AI. however, you want without requiring any sort of standardization. but the most important part is you have to be incredibly thoughtful, about how your engineering team actually uses the systems under the hood. when we started building bammo, i didn't think it would be possible to build some of the software we did. and just yesterday, one of my engineers built a partial c compiler purely in bammo. so when you start pushing the boundaries of these systems and you stop reading the code in some ways. in my mind, it releases the flood gates for engineering team to actually cover the gaps of what existed in your old process.

Speaker 6: have you ever worked at a company that had no CICD? they said, adding CICD would slow us down. they did slow down for three months while they added it. but after that, they move a lot faster. our processes have to evolve if we are going to ship at agent speed. and remember, this is the least amount of slop, your code base will ever have to this day. so just embrace it and start fighting it back. i fell in love with software about fifteen years ago, and it was the first thing that truly changed the way i perceived the world. and i really genuinely don't want software to die, and i think we can all build a world of beautiful software. and i think what it takes is i want each of you to go home today and build these sloppy tools, make your systems more robust, make your processes more robust. and then for the bravest of you, i want you to go back and think about these core foundation layer systems, think about how they're broken and see if you can imagine a way to fix them. i think we do need a new git. i think we do need a new database. and yes, i think we need a new programming language. i'm vivek and i work on bammo. thank you.

Speaker 7: it was sham gupta, and i was offering delegate meter, a king on building of training and inference infrastructure for the matter, supportalab and infrastructure organization. today, we're going to be talking about production of agentic systems. when most people hear the word evaluation, they think about benchmarks a model scores ninety percent on benchmark, a new version, scores ninety one percent, a team celebrates a little bit. but agentic systems have fundamentally changed what evaluation means. today, the system stores the prompts, the retries, the plan, the tool calls, the tool information, the execute workflows. they interact with the production infrastructure. the question is no longer did the model generate the right, right character. the question is the system behavior quality. do they?

Speaker 7: I would like to discuss how evaluation is evolving for modern range marking into blution infrastructure. This is the problem. Almost every AI organization is encountering today. Offline benchmarks continue improving, yet production reliability often remains unpredictable. Why is that? Because benchmarks measure model capability, production measures system behavior. A benchmark doesn't capture dull failure, AB test changes, use availability, long learning growth lows, and as systems become more dynamic, the gap between the benchmark performance and production performance grows. The result is what many teams experience today: high benchmark scores, but ununderlove the production behavior. Traditional evaluation focus on outputs, but we should ask a question to the model: did it produce a guide answer? Agency systems force us to ask a different question: did the system behavior and dedupllanning quality, tool usage, execution, recovery from failures, decision making? In other words, we are moving from evaluating answers to evaluating architectures, and that requires fundamentally different evaluation architectures.

Speaker 7: Many teams still think hallucinations are the primary failure modes in production. They are often just one category. Agent systems introduce a whole hierarchy of failure modes. At the very foundation, there are memory failures, repeatable failures, safety failures. As you go up, you have to think about reasoning mistakes. People banning rily ilecuand at the highest have to think about coordination failures. And this is why evaluating only model output misses the most production rescue observed. While the most useful mindset shift is to stop thinking like researchers and start thinking like SRE or approaching SRE. Don't measure success using accuracy. Measure reliability, latency, cost, and agent systems take the same approach. The goal is not maximizing the benchmark scores. The goal is to maximize dependable outcomes. Reliability becomes the north metric, accuracy becomes only one boot. And this grammar is how I think about modern evaluation systems.

Speaker 7: You can see they are benchmarks. They are useful, they're scalable, they're repeatable, but the operation range is limited in the middle. The scenario-based evaluations, these simulate realistic workflows. And the way to see promotion delimilitary. This is where the highest value evaluation signals come from. The surprising insight is that the most evaluation data often comes from real users interacting with real systems. Now let's talk about, it often involves so often evaluations still focus on much dollar changes instead of evaluating prompts themselves. For example, a customer support flow is a coordination flow. A search flow operates inside the immediate environment. We measure the condition accuracy, tool selection, planning quality, resource usage, which becomes exponentially high at high scale. The key technical takeaway is that evaluation should be signal-driven, not flow-driven. Once a system hits production, every interaction becomes a signal. This is one of the biggest shifts in evaluation thinking. Production traffic is no longer just traffic, it becomes evaluation data. We connect execution traces, user outcomes, escalations, failures, feedback signals. Production is the largest and the most representative evaluation data any organization will ever have.

Speaker 7: But how many organizations view humans as feedback systems? I think that's the wrong framing. Humans are the evaluators. They provide signals that automated systems cannot: subjectivity, trust, usefulness, and safety. These signals become really critical for calibrating evaluation pipelines and identifying blind spots in automated metrics. The most successful systems combined automated evaluation with targeted human review. Now, agent systems drift constantly: model changes, new version every couple of months.

Speaker 5: The prompts can change, tools can change, user behavior can change.

Speaker 8: I'm co-founder of a company called Human Layer. And I'm here to talk about loops. I think we've all been building loops lately. And I realized recently, I think we're all doing it wrong. Loops are really powerful. Don't get me wrong, but so much of the discourse around them is hype-driven and just really not helpful, right? I think we have this idea just kind of as an industry somehow that we can like pipe a prompt in a loop to a coding agent and that we can build things this way. Maybe we're investing a lot of time in verifiers. Maybe you have six different code review agents. But at the end of the day for doing this, we're still building forty thousand line prompts that just nobody wants to read, right? And this isn't exaggerated. Jeff Dean, right? This is real and it's innovative. Um, it's a sharp tool that works very well for certain types of problems. It works very well if you're not going on a team, it works very well if you're not working on critical systems. But most of us work on critical systems, and we don't fit in that box. So today, I want to talk about how to build loops that work in large, complex code bases for systems that have real customers, real users, real regulatory obligations and service level agreements and everything else that keeps us from shipping yellow forty thousand line prompts straight to production. In other words, I want to talk about how to build loops for the real world.

Speaker 8: If you're not aware, uh, this post actually dates back to July. It went viral this past January, which is when a lot of us, I think, started building loops. And of course, much more recently, I'm sure you're going to see this slide a lot this week. But Peter Seibel said that we shouldn't be prompting coding agents anymore, right? We should just be designing loops that prompt agents. OpenAI notoriously built on loops. They built loops, built code, loops reviewed the code. They logged in, released the code. They found and fixed the code. There's even loops for finding and fixing bugs in the loops that are merging the things, right. It's loops all the way down. It's a loop journey. Also the Creative Clock code recently said that this is his entire job as an engineer. Now it's just writing loops to prompt Claude, and eventually we might not even need loops, right? We're just going to have like swarms of agents designing loops, prompting agents building code for loops. And like I don't know somewhere, we're like writing production code. I assume and in fact, the loops that we're building are producing so much code that we can't possibly read all of it, right? So we might as well just not read any of it, right? We're investing in verification and code review, but all this code is read only. This was the thesis of a conference that was here in town last month. So a lot of smart people at frontier labs think that this is the future of software development. And if you're doing this, you're moving much faster and everybody else is getting left behind. Now, it's not clear how well this works yet. UIT took six months to fix the Claude code terminal flicker. The OpenAI team wrote a renderer in a fraction of that time and OpenAI's code notoriously has stability issues. What is abundantly clear, however, is that this stuff is really expensive. If you don't work at a frontier lab and have an unlimited token budget, all this code that we're writing is actually really expensive, right? MD Pocock talked about this recently: bad code is much more expensive in the age of agents than it has ever been at any point in the past. So today, I want to talk about what I think works in the real world and what we started doing at Human Layer, which to be clear, is still building loops, right? I think loops are super powerful, but I can design loops and still read the code. In fact, we can design loops that make it easier to read the code because the loops are making the code better. We can solve hard problems in complex code bases with loops, and we can build our software for that at three incrementally. But to do this is going to take some real engineering, yeah.

Speaker 8: So let's talk about control theory. Control theory is all about how we drive a dynamic system, which should be your code base, towards some desired, stable or optimal state, right? You have a sensor that measures the current state of the world. You have your set point, right? The desired state of the world. The difference between these two things is measured error. You have a controller that reads that measured error and turns it into a control signal about an incremental change to be applied to the system. We have an actuator that applies that change to the system, which is undergoing disturbances in the meantime, and then we remeasure, recompute our measured error and we go back to where we started. Now this sounds really complicated, and it can be. I have a twin brother, actually, who's an aerospace engineer. This is how they keep fighter jets from falling out of the sky.

Speaker 8: ah HH.

Speaker 8: it's probably a little bit simple than most.

Speaker 8: you'll think does that know what happen of these a thermst stat uses contcontrol PP for for any our european friends in the audience？

Speaker 8: that's part of something we have here in the states.

Speaker 8: it's called air conditioning.

Speaker 8: and uh, most of us probably actually use control loops on a daily basis, right？

Speaker 8: cuberonetties auto scaling systems are built on control loops, infrastructures code uses a desired state current state iterative changes like control l oppattern postgresses are autovacuum and rex iritrol on both both uare are proproed imate control loops.

Speaker 8: 是的, 好的, control loops are ideal when we have a system that we want to change a problem. we can measure in a way to get feedback on the result of that change, like good soft engineers have always been thought to do control, loops, changes, system incrementally, instead of just trying to get straight to the end state immediately. all the once and risk blowing everything up, right？ they help us to avoid over stistering and destabilizing the system. and adminimizes risk.

Speaker 8: so control loops will the opposite of what i'm going to call our blind rall flu. they're how we avoid pours that look like this because nobody wants to review this right, which is not to say that all rouf loops are blind loops. the best roops are actually applying control theory. i know jeff hunleis out in the whole somewhere wandering around. if you go talk to him, he's going to tell you the same thing, right？ right？ the routh is, is teaching device. and i think some of us ready it a little too literally. but this is how we should have always been doating loops, but it's other issue e the rouf loops. they're not incremental right？ it's just a bash loop.

Speaker 8: so we have to build a gentic control loops. and to do that, we start by defining a set point, which is the desired instate of our code base with respects to some property of it.

Speaker 0: and we had a sensor. there's a lot of ways to build a sensor. it can be strictly deterministic. your ES land rules, your AST grap, your packwork or it can be not deterministic. you can have an agent and a skill and a bunch of natural language rules. so you could also just have a pipeline like a combination of the two. so how do we build a gentic and goops？

Speaker 0: 嗯, there we go now ah this is all theory, right？ practically speaking. and because we're using agents, we can bw the lines a little bit between system components. so eight and biys react doctor, for example, is fantastic. it is um it's a great way to catch all of the react slot that lalad snot ck into your code bed last week. but it's a hybrid sensor in controller. it tells you what all the problems with your react code. and also, by the way, what of the top three things you should fix and how how do you fix them？ similarly or controller and actuator might actually just be a single agent deciding on an incremental change to make and then applying it and the same context window.

Speaker 0: but i want to suyou in on the controller a little bit, because without one or without a well tuned one, we might make too large of a change all the once, or we might might make the wrong change entirely. and if you put that in a loop year in trouble pretty quickly.

Speaker 2: 嗯, 对对.

Speaker 0: so we can use control loops to root out bad patterns into clean up our code, but we can actually use them for all sorts of things. right？ we could make sure that our api is complained with someone else as open AP speback. we can make sure that our mcp server is complained with whatever version of the the mcp specification that were currently on. haven't checked. you could mirr a project from pyython and hydescript or vice versa. you could even maintain your veet based slolop fork of next js against the upstream. the key questions are. can we find something we can measure？ can we apply changes incrementally？ and can we get feedback on the quality of those changes to illustrate that i'm going to walk through a control looloop, but we use eternally a human layer for our loop.

Speaker 0: we are incrementally migrating our RKCAPI to affect. we adopted it for some of our race prone code. we like it so weadopting in across the rest of our code base. if you've never seen effects code before the code on the right, just just kind kind tritrivial proycure, the left ft rewritten effeffect h. this in tax is really weird were cychoes. we really like it. it's not for everybody. that's okay. uh isn't n't talk about effect, so will keep moving wh clickers not work king cool.

Speaker 0: so step one, we have to build our sensor to find unmigrated procedures. we can have an agent do this or we could use grap or grap grap. but in suwewan can use ST grap because it's really powerful. it's a tool to have in your tool box for building loops. it's language agagnostic. it's out of band from your touchtript config or eelluent rules, which if you're a touchgroup developer, you have watched lalaud disled those with inline comments, but so we can just write a simple rule that finds on migrated procedures res based on the pattern above.

Speaker 0: and we, over time, we can even layer on more rules that describe other patterns. we want to get rid of with granular, including exclude pads. if you have a multi lingual monorepoo, like we do i'd work for any language you could possibly imagine. and we 're just sandour code bastions and it'll ce ce, a longest of violations. it'll too long. in fact, it'll give you about fifty liys per violations, and we're just going to filter it down to four, and we're going to sort it deterministically. why are we doing that？ at the beginning？ i said this was going to be practical. and so we're going to step outside of our control of paradime for a second. because before we start incrementally migrating procedures wanted a time, we need to enforce that all new procedures are using effect, right？ so we're going to run a full scan once on main, sort all the violations deterministically and tracit in our verversion control.

Speaker 0: and then on every new PR, we can see if it the branch added any unmigrated procedures, right？ so this is our control loop, and our system is undergoing disturbances. in this case, uh, our temage shipping clothslog. and this is how we make sure that they're not undoing our loops work. this doesn't matp directly ads to a part of the control loop, but we can kind of like squen edit a little it. it calcalled a disturbance dampener.

Speaker 0: so now we've stostothe the leeding, we can actually design our controller for a simple controller. we could just deterministically pick the first violation from the list. you can use bash and jq, or we could get a little clever er and use ST gt to find the smallest on migrated procedure and always pick the smallest one to reduce the risk.

Speaker 4: 嗯.

Speaker 0: we could have an agent make the decision. if we really want to, i don't think you should ever send an agent to do deterministic codes 's job, but you certainly can, in fact, depending on the complexity, we could have the agent, take the procedure to migrate and just do it at the same time, like we just talked about, but we can make this even more powerful, right？ because we're not just migrating effect for the sake of it, we're doing it because it's helpful for handling errors and for helpful us instrument our code better.

Speaker 0: and so what we could do, if we want to get really clever is we can look at our telemetry and figure out which procedures have the most errors or or the least instrumentation or has a gap in our APM, right？ and when we send a control signal to our actuator agent, we can include not just the procedure of migrate, but also all the data about the things that were trying to fix with this migration, so that the actuator agent can actually make the code better instead of just doing a one to one migration. um m. there very go.

Speaker 0: so next is building our actuator. our actuator is just in agent, plus a skale brbring your sel, i coding agent of choice. you should spend a lot of time on the skill. not all that should be at front. you'll want to iterate on it over time based on what works it human layer. we like to build up what we call golden patterns by hand before setting the agent lose. these are just like idetomatic hand written examples for the agent to follow because they're just pattern replicators and otherwise you're getting what's in the dogs or what the agent knows from the internet.

Speaker 0: and so we pipe the skill. 对对对对, plus our control signal and do our actuator agent. and if the skill, of course, should include a response template and the agents going to work and work and work, and it will produce a final response. and then we're going to deterministically commit and push and create APR using the final message as rpr description.

Speaker 0: now all we have to do is actually run the loop, right？ um my recommendis is you get get acactions your your gt lab or your circle ci or whatever else you're using, because it has access to your code. it has access to your secrets, and it has great dispatched in scheduling primitives, right？

Speaker 0: we don't need a new cluster for this. so we can write a workflow that runs a single iteration of the loop sense, control actuate and creates a PR. and then we can schedule this to run once a day. and every morning, we walk into the office to a small incremental PR. that's a risk. and when we first tried this, this was actually really frustrating and we turned the loop off. and um because we had to constantly update the skill, we had to constantly check out the branch, changed the skill, changed the code, commit and push, and our loop is actually really high friction, right? but there's a better way to do this, right, where we can put a human on the loop in a really low friction way to resteer it when it goes wrong. and the way to do this is to just create a feedback file that's tracked in version control, just as a markdown file, right? we can deterministically load it into our actuator agent's context. every time that it runs after we run the controller, then we can add a label to the PR, right? each workflow needs to be able to identify PRs that it created since there'll be a bunch of different loops running. and we want one workflow to respond to feedback from comments on their PRs, and we're going to have a comment trigger to each loop workflow. so when a user leaves a slash iterate comment on the PR, the loop workflow is going to pick that up. it's going to deterministically load all of the PR context, the diff comments, the comments, the description into the agent's context, along with the skill, and it's going to instruct the agent to fix the code, but also to update that feedback file right looks kind of like this. and the benefit of doing this way is that now that feedback file with instructions, it's backing in your version control, you can see how you've changed it over time. you can revert it if you need to.

Speaker 0: um, so the next thing we're going to do is add flow control because the other problem that we had when we did this was that if we were at a customer site for a week or if we were traveling or spent six days working on slides instead of writing code, uh, the PRs from all our loops, they just stack up, duplicate work that conflict, and we couldn't get around to doing it. and like the loops work is important, but it's not that important. and so now we just had all this like junk, we had to deal with it. it wasn't important. so this is actually really easy problem to fix, because each loop and its workflow has a label. they get attached to PRs when the workflow first runs. before we check out the code and install the dependencies and run sense, control, actuate steps, we can just check and see if the last PR that we created or any PR with the loops label on it is open. and if so, we just shut down right, because this means that the last time that a human reviewed the code from this loop was before the loop branch, right? no human reviewed the last output. so there's no reason to stack up even more work for humans to review. right, right, right, this way, we have exactly one PR and most open per loop at a time. no stacking, no duplication, hopefully no conflicts.

Speaker 0: and of course, once we're feeling confident in the loop, we're going to want to speed it up, right? i have one hundred and fifty RPC procedures to migrate. if i do one at a time, it's going to take six months, which is way longer than i want to wait. fortunately, there's a lot of ways to pick up the velocity of our PRs. we could have our controller pick procedures to migrate instead of one. there's a lot of ways to migrate. we could have our controller pick three or five, and then do each of those in a separate implementation phase will be both cheaper and more reliable, since each migration gets its own context window. or we could just run the workflow four times and give one PR to each of four people on the team. so let's put it all together. we build a control loop that improves our code incrementally and we're actually reviewing the code. it has adaptive flow control. so we're not creating a bunch of fluff or a bunch of work that nobody wants to review. and we can resteer it on the fly in a super low friction way. if you want to try this yourself, um we built a skill, please try it out. my twitter handle is down there on the bottom, please share it. i would love to see what you build. and if you get excited by this, a human layer, we're hiring here in san francisco. and if you're working on mission critical systems and want to figure out how to get more out of the AI, we'd love to chat. thank you so much.

Speaker 8: um, in twenty twenty six, coding agents will quietly retire the first software platform, not because it's bad simply because the platform is unnecessary. i am dominic tunnel. i am founder and ceo of resonate. resonate is a global execution platform built with minimalism and simplicity as its core technical values, and these properties will play a central role in this talk. at resonate, we have a belief that software engineering is being disrupted. general purpose implementations will increasingly be replaced by bespoke implementations, generated on demand, not as a new library, a new framework or new platform. but as a minimal extension of the infrastructure that is already in place. if this theory holds, tools will move up the stack. instead of reusing a general purpose implementation, we will reuse a specification, and we will derive a bespoke implementation from it. in fact, we can build many bespoke implementations tailor made for the infrastructure that is already in place. we just have to ask the agent at this point. the prompt is a platform. resonate is a global execution platform. we have an implementation of the resonate java. we have implementations of the resonate sdk for typescript, python, rust, go and java. so we have to ask, what does this new reality mean for us if implementations become generatable? where does our value live? and our answer, our value moves from implementation to specification. now this changes how we think about reasoning. the product is no longer an implementation. the product is specification, the protocol. and from that protocol, we want to derive multiple server implementations. one is a general purpose, resonate server. our reference implementation. others are implementations built with infrastructure partners for customers and partners. this means durable execution right on top of their existing infrastructure with minimal additional dependencies. so the question is no longer can we build a server? the question is, can we repeatedly synthesize trusted servers from the same specification? and if so, how? when we talk about agentic engineering, we focus all of our attention on verification. how do we know the result is correct? but today, i want to focus on the specification instead. and more importantly, how can agents participate in specifying the system, not just building or verifying it? now recently we partnered with multiple infrastructure providers to bring durable execution natively to their technology stack. one of them is syadia, the company behind nats, an open source messaging system designed for building modern distributed systems. for the rest of this presentation, we will use resonate on nats to explore our agentic engineering practices. how do we go from specification to implementation? first, we need to level set our mental model. this picture is a common view of agent coding. there's an agent, there's a specification, and then there's an implementation. and for many applications that is enough, but it is not enough for what we are trying to do because we are not trying to generate one implementation from a specification. we are trying to generate multiple target specific implementations from the specification. so the specification must not take any aspect of an implementation into account. the specification must not assume a concrete database schema or concrete indexes. this specification must not even assume a relational database with tables and transactions at all. it must not assume a key value store. it must not assume weak consistency. it must not assume strong consistency. the specification must be abstract, only the implementation must be concrete. so we ask the agent to follow the abstract specification and generate concrete implementation specifically. at first, we asked the agent to build a resonate for us in rust on top of foscers, and the agent failed. the gap between the abstract specification and the concrete implementation was too large. the agent generated a system that worked on the happy path. it passed the basic tests, but it was not correct. it broke on the concurrency. it broke on the process failure. it broke on the network failure. the implementation was closer to a prototype, but not a production system. so we amended the process. instead of asking the agent to jump directly from abstract spec to concrete implementation, we inserted an intermediary artifact, the concrete specification. that concrete specification was derived interactively with the agent.

Speaker 8: but the human was the main driver for postcores that meant making target specific decisions, explicit the data, schema the indices, the SKL clearies, the tracacboundaries one. so those decisions were written down, the agcient was indeed able to implemena production system. so this worked, but it also revealed the limitations. the agent helped us build the system, but the agent did not help us design the system. and if the specification is a reusable product, then that's not enough. now the next step is obvious. agents have to move up stream, but how when we started building resonate on that sale, we changed question. we did not ask can the agent built the production system? instead, we ask what does the agent need in order to design the system first and build the system. second, so we gave the agent access to a deterministic simulation environment, and we gave it a different task. do not build the production system, build a simulated implementation. the simulated implementation is not the product, it is executable design its purposes to discover the correct algorithm under partial order under partial failure. and once these algorithm are discovered, tested and verified in simulation, then we ask the agent to write the concrete specification, and only then to be ask the agent to write the production implementation. so the process becomes abstract specification, simulation implementation, concrete specification and then concrete implementation. this is a point where the agent moves upstream humans are still involved in the design process. but now the agent is a driver, two ingredients make this possible minimualism and simplicity. unfortunately, minimualism and simplicity are not the starting point. they are the finish line wespspbeen years years, making proprococol smaller and simpler. every time we ran into a problem. we ask, what can we take away, what abstraction can be race, what property can be remove, what relationship can be break. the result is a very small protocol sent that around two objects, a global promise and a global task that simplicity matters, because even simple concurrent disssibuity protocol have a complex state ment behavior space. so another terms implementing even simple protocols on top of a few simple primiters is tough.

Speaker 8: let's make this concrete with nuts. nuts gives us a small set of primitives. we can build on quees, a key value store and delayed or scheduled messages. these are not reasoning concepts. these are the concept of the target platform. so the design question becomes how can we express the reasontly protocol using only these primatters? let's focus on the key value store. the key value store is versened. we create a key with value fool, then be updated to bar, then be upted it bbus. so the latest value is pass at version, too. most of the time when we read the key, that is exactly what we get a fresh read. and if all reads were fresh, the design will be straight forward, but sometimes the readed is stale here. lalatest value still bbus adversion too, but the read returns full at version zero that is not corruption. that is not a bug in the key value store. that is a valid read under the consistency model of the target platform. and that matters because our implementation cannot be correct. only when the target behaves conveniently, the implementation has to be correct when the target behaves legally. so the simulation environment has to expose exactly this kind of behavior, fresh reads, stale reads. and the version information that tells us which word we are in. unfortunately, we do not know the read was stale simply by reading. we will find out later when we try to write here to we read version zero. so we try to update version zero, but the key has already moved on the right fares. that is the moment the targetels us. the word you saw is not the current word building, always correct applications on top of a concurrency model that allows occasional stale readis, not simple, not for humans, not for agents. so how to we set up our agent for success, what tools does our agent need to acist task instead of falling flat on its face. agents thrive on feedback, immediate unambiguous feedback, not just feedback that shows this man wrong feedback that shows why and how this man wrong, what state value was returned, what logic was triggered, what ride failed and which invarian broke because of that. so we built a determiniisc simulation testing environment in python. and inside that environment, we simulated the parts of nuts. io. we depend on here, for example, is the simulated key value store. it keeps a full version history for every key on get the simulated storard, sometimes return the latest version, but sometimes controlled by the deterministic random generator. the store returns on lot diversion on update the store and forces optimistic concurrency the right only succeeds. if the version you read is still the latest merger, otherwise it raises. this gives the agent a store that behaves like the real store in the ways of metaphor correctness. but unlike the real target, the simulation 's deterministic, it's repeatable and it's inspectable. so when the agent writes the wrong algorithm, we can reproduce the exexcution that broken and the agent can repair the algorithm against that trace. but deterministic simulation does more than just inject state rates. it lets us expose facts. the real platform heights. we call this the forbidden fruit in production. when you read from the key value store, you only get the value in the version you observed. you do not get to know whether that read was fresh or stale. you do not to see the latest value. you, you misand you you shouldn't get that information, because real code cannot depend on it, but in simulation, we can recorded here every get in mits a tray event. if the read is fresh, the tray says this was fresh. if the read is still, the trace says, this was still this is what you got. and this is what the latest value was, that information is forbidden to the algorithm, but it is incredibly useful to the agent. it lets us explore failures in terms. the agent can act on now.

Speaker 8: this is what a tracevent looks like. the production code only receives the result. it sees the promise was pending. this is all the real platform bod tallas, but the simulation also records the type of the read. this read was a stil read, and it records the latest value that was hidden from the algorithm. the latest value says the same promise is already settled. that difference is exactly the kind of facct and agent needs when it's debugging, a distributed algorithm, not just the invion field, but the invion failed because alrithon made a decision from a stale view of the word. again, the algorithm is not allowed to depend on this information, but the agent is allowed to use it to explain why the algoriththm it designed was wrong, cause and effect becomes visible. the agent does not just learn that the system is wrong. it learns why the system is wrong. and with this approach, the agent was able to close the gap. first, the agent boot, a proof of concept in the deterministic simulator verified by foce testing from the proof of concept, the agent derived the concrete specification. when we already knew that irorithm was correct. and like before from the concrete specification, the agent derived an implementation deterministic simulation, lets agents participate in the design, not just in the implementation. humans are still in the design process. but this time, the agent is a driver from a single abstract specification, the agent designed and builook the platform via simulation to concrete specification to concrete implementation. the prompt is a platform, and the specification is a product. thank you very much for watching. if you have any questions, please don't hesitate to reach out. you will find me and resonates this code.

Speaker 10: hey, everyright. i'm swim. i'm i'm see you in good fronend of stulid search. and today, my talk me ter use a signals die at retrieal boundary. so look into uh what are agents? essentially, my agent feels what is uh cause of fiels in retrieval, particularly and how to make actually signals ross, the trivial boundary and how to make your agent basically unconfair. 是的没的。 so, uh plus started ted. what is n't an agent? and agent is an analym that has agency to reason involve those. uh interact with that, dual would retrieve uh the memory to complete. the ask, ask major or uopa is missing is a learning. it should also learn from what to work and what it didn't work. suppose um if i have to uh explain, what is asient did i can explain through occasient. so if i have to explain uh, what agent does, i had explained with react agent. so basically uit from the agent uuh execute it in a look,

Speaker 10: uh cling to a retrial research and then pause when the the do's this company.

Speaker 10: this is very basic react architecture.

Speaker 10: one thing that is missing is how to make agent the noledge on the outcome.

Speaker 10: so asian keeps feeling at the same task gotten the report, said eighty five percent of air takes failing chin, concerts and mkenzees to entwenty deflight report.

Speaker 10: the problem came out to be most of the time is that the tually was static, seventy three percent because of tratrict and conacst from from from around ving.

Speaker 10: for the wrong thing h.

Speaker 10: we have been optimisze for the wrong thing.

Speaker 10: you are paying and not for your potimisong.

Speaker 10: and we have been optimising for the wrong thing.

Speaker 6: the 嗯, 没有啦啦啦啦 the 我是你 the 嗯, 嗯嗯, 没有 the 我会你.

Speaker 11: ladies and gentlemen, please put your hands together and welcome back our mc member of the technical staff at keycard alii.

Speaker 6: how 嗯.

Speaker 12: welcome back to our afternoon. keynotes, i hope you all had an amazing day so far. i know i certainly have. i've enjoyed exploring the different talks across all of our different eighteen different tracks. i've enjoyed exploring the exposession, and i'm enjoyed talking and at working with all of you, a engineers, a really special place where we can all come together in one space and have really interesting conversations that pushed the boundaries of what snacks. it's a really powerful thing to get AI security engineers. any engineers into the same room and understand how we can of unlock um shipping because of security issues, ship faster and increase our engineering velocity. it's also really exciting to get the same people that are building the models in the same room with the people that are using them. so we can understand what actually works and help us to find what snacks. i'm super excited for the rest of our content today. but i want to take a moment to think our amazing sponsors that make all of this possible. let's take up run of applause for our presenting sponsor microsoft. i also want to recognize our lab and platnum sponsors also our gold sponsors, rn of a pause for them as well. and last, but not least, i want to thank our silver and brown sponsors,

Speaker 4: which there are newable amount of those as well as eararling takes a lot of people to put on this event.

Speaker 12: and so super thankful to all over their responsors thanful for all of the people that are here that make this community possible. um i know weve heard a lot about offacfactories today, and they're super interesting weall. all very, very alert by the promise of what the productivity and unlock that they can deliver. but in practice, they could be more difficult than in theory, and i'm really excited because our next round of speakers are going help us understand the practices for a loop engineering and hell, ll us understand how we can create software factories that actually work and don't produce slab. i'm super excited to introduced our very first speaker of the afternoon keynote sessions. um this speaker is a seasoned ai speaker. he to told me is his fourth time you spopoken at a engineer, which is incredible. um this speaker is known for raaging. the war on slop. he is known for calling the the term context engineering, and he is known for creating the research to plan implement framework that we know in love. it's my honor to introduce jackfororthees co founder and ceo. he were layer to talk about harness engineering is not enough. why software factories fail? please welcome to the stage deck.

Speaker 6: 我开了你 for very money are was doing guys giit up for all the great speakers today.

Speaker 13: so far, all right, this is hardest. engineering is not enough and why software factories fail and we're going to click.

Speaker 6: maybe um oh.

Speaker 13: that's way too many slides hold on guys. okay. so we're all racing to put eye coding into production. and there's been lots been said about loop engineering, and we should probably write more loops. and yeah, II, guess we doing loops? now strong gm, built a lights off software are factory where nobody even raiss the code. and the prevailing narrative is we should just end more tokens. you are the bottm eck. the models are good enough code is free, just ship more stuff. but at the same time, we are starting to see the cracks, our friend morio at ai engineer, europe begduus to slow down because companies that should not be having outages because of coding agents or having outages due to coding agent, perhaps um code bases are falling apart faster than they ever have before. and our friends and far, as say, i actually even did a report since we all picked up all these AI coding tools in january, a be february ary um or request coreview quality is wait down, were having more comments longer comments and tons of ps being worged without any review at all incidents are way up. bugs per developer a way up, and many people will tell you that you're holding it wrong. that's the only reason you're not. well, maybe you are, but that's not the point, but i've spoken a lot about how to hold it better when it comes to working with AI, probably ly million views on youtube, but this point across a monunch different talks. ks. and the basic thing is like, as engineers, we've been told that if token maxzine isn't working than it's a skill issue, you just need to spend more tokens um with enough reading the code that with enough harnest engineering. if we maybe sprinkle some magic words, adversiial review, enenough or PR bots that we can get the best of both worlds, ten do one hundred exfaster high quality. and nobody has to do that thing. we all hate called code review. i'm here to convince you today that this is, in fact, not a skill issue that that no amount harharness engineering or loops maxine can solve what is fundamentally a model training issue. that's why we say the hardest is not enough um and to understand this, we kind of have to grapple and digin to how coding models are trend. i'm going to talk about what i think the shortcomings are with some of the current benchmarks and what better ones might look like. and it's will talk about how to move faster safely in the meantime, um it's because sound like a rant um but is hope here um i'm going to talk about our journey and a bunch of the land minds we've hit um building in this world, a bunch of exciting new techniques that we've been king king with a lot of our users and customers dedevelop. and i think how we all as a community get to the next chapter of agentic engineering after whatever this thing, if that were in. um so we use a lot of words here. i assume out a little bit. i want to give you kind of like a brief history of the software factory. um and if actually, i don't i just learn this last week is the term software are factory was defined a nato conference in ninenineteen sixty years. you'll start around twenty twenty two right right before II started coming around um and basically in a typical twenty twenty two software factory. you will have some people building stuff. youll have engineers, you'll have pms. maybe you have some sort of leadership teen that is driving the vision here, and they all the side that stuff needs to get done. and so you put it in a tracker, a linear, a gera abeds, some sort of state machine, the tracks, what needs to be done. and then someone goes and grabs something off for there. and they build the thing. and there maybe some automated testing in that process, mabe be might manual testing in that process. and a certain point we make this poll request thing says, okay, cool. we're got to run about to checks automated stuff, a humans going to review the change and review the code. and perhaps we might even have uh a human pull down and test it somehow. and if anything goes wrong here, we look back to someone builds the thing and eventually, we're ready for product. and so we shift it to production. and once it's in product makes contact with our users and users, do a thing that we all love. um users love to complain. i love our users, but yeah, they're going ask for things. things are going to find bugs, you're going to file feature requests um and that goes back to your team. you might also add monitoring. and so um you know what do we want more than anything se? and we want to wake up engineers at three in the morning when something breaks. so then you get dragged out a bug to try to go fix it. um we waon on on in this loode. um and we am ship a bunch a code. um and one thing we we notice here is that teams figured this out decades ago is that this someone builds the thing step 嗯 is usually going to take hours or days in most cases, and the review part will also take hours or days for large things. and so team started doing these upfront planning, architectural proposals, sprint planning, and it would collaborate on on on these things as a team, uh, with the hope that we might decrease the percent chance that something would need to be reworked that we would be able to reduce the time spent in reviewing every line of code because we aligned on everything ahead of time.

Speaker 13: this brings us to the agentics software factory every company and their mother is talking about how they built a coding agent factory that shift seventy five percent of their code.

Speaker 13: now, literally everybody.

Speaker 13: and so we look at the software factory from twenty twenty two. We just replace someone builds the thing with an agent builds the thing. And we have an orchestration in a harness of sandbox in a model and computer use. And i'm not going to get into the details of that. You can watch one hundred talks about that this week, i'm sure.

Speaker 13: but now the building part takes minutes or hours, but this human part still takes hours or days if you're going to review the code and you're going to test the changes. And so we bring an agentic code review, and we bring an agentic regression testing and it makes this part faster, but it's probably still the bottleneck, but we can do more loops here. Why not? Let's do some more loops.

Speaker 13: so we can route all incident in straight into the factory. So does someone need to get woke up? And try to fix it when they try just wake up to a power query. And maybe that fixed the issue for you, you can take all the user feedback. So just stick a straight into the factory. So that people can ask for stuff, and it gets built. And now your only job is how much things can you stuff into the queue of stuff to do and how fast can you review and test the changes.

Speaker 13: which brings us, of course, to the lights off software factory, where basically we no longer read the code. We say, you know what this is going great, that code review thing. No, thanks, or it is not going to do that anymore. And we invest into all these other parts of the system. You're testing your monitoring, you roll out everything else. We just write more code and build those systems better. And now our job really is just how much stuff can we ask the agent to build.

Speaker 13: i am going to posit that this does not work. And this is why software factories fail. As i decide what i'm going to say, has nothing to do with Willem Larsen had this great post. I'm going to literally take his quote for beta. A developer providing coding a side project a dozen people will ever run, and a team keeping a ten year old enterprise system alive for another quarter share, almost no constraints worth naming. And most of what you hear on the internet is one of these groups of people telling the other group of people how to live their lives. So if you love vibe coding, please go on. What we care about is, how do we help people solve hard problems in complex code bases?

Speaker 13: we are brown field a lot, which historically has meant like some ten year old job a thing. I actually think agents really start to struggle after maybe three to six months, especially with the pace at which we can ship. Now you can ask me how i know this, and i will tell you that is because of Joy, old, twenty twenty five. We tried this. We went full lights off. If you have tried this seriously for number of months, you probably found at least one issue that the agent couldn't solve. Even with your most advanced prompting, you do research with do reproductions, you just you have to go and dig into that code base that you stop reading three months ago to try to figure out what's broken. And in the meantime, your site was down your users were pissed. And you, if you were like me, you were probably miserable reading all this slop code that you let slip into your system.

Speaker 13: and what i want to get to is basically, models have a shortcoming. They can't maintain an improve code base quality over time, not without a decent amount of human steering. And i want say, maintainability basically talking about issues like it becomes really, really hard to make a change in one part of the code base without breaking other parts of the code base. This is Martin Fowler, low shock on surgery, textbook code smell. There's a bunch of books even go read about it. In fact, John Ousterhout, who is actually here speaking this week. So you can go ask him in person about the philosophy of software design, if you want to.

Speaker 13: but it brings us in this question of, like, why can't models do software maintainability? And you may also be saying, but dex, you know, surely the models have gotten much better since then. They've got better in some ways, but they still about the same in others. If you want to solve one of problems or vibe code, a new marketing site. Yes, they got way better since twenty twenty five and twenty twenty four. But as far as improving code base quality i think they not got gotten much better. Now. I cannot prove this because there are no good benchmarks for a model's ability to maintain code base quality. And i'll getting it like where we're going with that. But if you work with coding agents for a while, a lot of people probably have this feeling that they generally make things worse over time and make the code is harder to work in.

Speaker 13: to figure out why this happens going to zoom out to the first great coding agent. Why did Claude Code go from nothing to four billion? And i think now they're nine billion in revenue in under a year because there were great coding agents before Claude Code. You had Aider, you had CodeBuff. There was a bunch of tools in this category. They had all the same tools. So what was the difference? The difference was that, that this was the first time that a model lab trained a model against the harness that they were going to distribute to users in. And it got really, really, really good. This is just some of the tools, but it got really, really good at calling these sorts of tools in an agentic loop. In fact, the OpenAI team did a talk in november about basically if you are a harness builder and you don't own the model weights, and you can't train the model in your harness. You will always be at a disadvantage compared to somebody who owns both the model and the harness.

Speaker 13: and i'm going to cite a couple slides from my buddy Calvin, French-Owen, who was a MTS on Codex during the initial launch. But models are just next token predictors. This is slide from over a year ago, where, basically, as you're doing your agentic loop context when it goes in next step comes out. And we've demonstrated this, but we're going to see we're doing coding agent and reinforcement learning in sixty seconds. So what we're going to do is get the model to get better at tool calling, better at solving software problems? We're going to generate a problem. We going to give it a problem. And we're going to generate a bunch of traces try to solve the problem bunch of different times. We're going to score them all on correctness and the test pass and all this stuff. And then we're going to reinforce, we're going to make the bad behavior less likely and we're going to update the weights to make the good behavior more likely.

Speaker 13: there's one of the classic here is SWE-bench multilingual. They're about fifteen minute tasks from open source repos like Rails and Jaco and Django and all this stuff, and they have binary one or zero rewards on. Did you fix the problem you were trying to fix? And did you do without breaking anything else? We look at actually your real problem from one of these benchmarks. This is Fastlane, which is a Ruby project.

Speaker 0: basically, there was some issue where we weren't checking for null. And we have a stack trace blow up because we have an null pointer exception. And in this benchmark, you have a base commit that we're going to check out before the issue was solved by a human in the past. We're going to give it a test patch that says, here's what the behavior should be. Afterwards. We have a golden patch. And these are hidden from the model. And so we have the agent go try to solve the and we store patch and we ignore all changes made to any test files because i'm sure you've seen models comment out test just to get things working. Because then we're going to apply our golden test patch. And we're going to run the test, all test and the new test pass and they both pass. Then we get the reward, otherwise we don't and the models are trying to get the test to pass. There's no way in this system that we can penalize it for poor program design or for eroding the maintainability of our systems. That's why we get things like this. Try catches around things that probably don't need a try catch or things like this. I think Biboga gave this example earlier of casting things, the other things just so the model can just wants to get the test to pass. And so you can't verify the maintainability of the code, it gets way harder to train on this stuff.

Speaker 0: So remember, picture verifying code quality, maintainability is orders of magnitude harder than the code runs and the test pass, because the cost function of bad architecture is measured in months and years.

Speaker 0: If you have a code episode and then you only find out months later that like somebody vibed this a little bit too hard, there's really hard to propagate that reward signal back across the gap.

Speaker 0: And now the first year's getting better slowly. And since I know someone's going in the YouTube comments about this, yes, I know benchmarks and verifiers are different, and they did actually have to be separate data sets, but they're shaped the same and the structure of these benchmarks directly correctly or fundamental.

Speaker 0: So, what is the future of evaluating code maintainability? There's a really cool one called SWE-bench Multimodal from Abundant AI? Or where you have like four hundred hour tasks of like clone all of Microsoft Excel, every single feature? Um, they have some sophisticated reward channel stuff.

Speaker 0: Um, SWE-bench from Data Curve is also like large tasks on OSS repos that were not actually in the training set because they were never actually built in the real world. Um, in the Frontier Code from Cognition, um, which is multi-PR tasks. They do interesting things like have the model write tests that if the model doesn't patch the code then it gets penalized. And we have a judge model that says, "Okay, because did this follow all of our code quality rules."

Speaker 0: Um, so we're getting better. But I think models judging quality can only go so far. Um, because if the new model, if the model knew what good code looks like, it would probably write it in the first place. Um, and review agents, during more tokens of the problem, it can raise the floor, but we're still constrained by what we can teach during RL.

Speaker 0: Um, and so I will posit for an AI, we're stuck reading the code, but we can move fast. And of course, there's a world where this is solved um in the future. And if you want to just keep throwing problems until you get to GPT-7 and want to think about this, by all means, please um, but I'd better less predict. We've got some problems to solve. So let's engineer our way out of this.

Speaker 0: Right, right, right. Um, so turning the lens back on. We're going to put the code review back. Um, we're going to embrace this approach of like how do we plan up front to reduce the chance that we have a long or a difficult review process? We're going to find leverage. We're going to use AI to help with this.

Speaker 0: Um, the first thing we're going to do is some sort of product review, understanding what problem we're solving. What's the desired behavior? Maybe looking at mock ups? Here's a product review as we were going on yesterday with a mock up of a new feature. Um, once we have our product review, by the way, we don't small stuff still goes straight to the agent. But once we have the product review, we will also do architecture, system architecture. A lot of people have been doing this for a while. Component contracts, data models, constraints.

Speaker 0: Um, this is an example of a doc that we build to understand how these systems are going to fit together. And what's like the high level picture of it from there? Then we do something that I think is really under-emphasized in coding these days, which is program design. I think people assume that once you get the architecture right, the model can just cook. Um, but we don't know, we often look into the types and the method signatures, the program layout and call stacks.

Speaker 0: And so here's some examples. I don't think you'd be able to read this one, but this is like the level of abstraction or rate is how we actually lay this stuff out and how these systems are going to interact. Um, Dylan M. from Cloudflare talks a lot about how he's using these call graphs as part of his planning process. I think this is exactly right. Um, and then once we've done the program design, we can do this thing called vertical slices, um, which is the order of implementation, multi-repo coordination. How we're going to build this across our entire system and how we're going to check it along the way.

Speaker 0: I actually have a bit about how models of horizontal plans and won't go too deep into it. If you want to learn more about this, you can go watch our talk from AI Engineer, Miami. Um, couple of shots of a doc like this, going through the tests and the steps in between each phase. Um, the main idea here is thirty minutes over here in pre-planning and alignment can save you hours in review.

Speaker 0: And so it's actually visible to still read every line of code. Um, skip this part. Basically, the summary here is like you don't have too many PRs. If you're drowning in PRs, you actually have too many bad PRs, um, because good PR is a joy to review. It's just really thorough. Yep. This is great. This is what we discussed. This is what we talked about.

Speaker 0: But even if AI PR needs twenty percent rework, which is generous for a lot of AI code flops, um, it's an emotional and intellectual burden on both the reviewer and the submitter. Right, right, right. Um, and so if you use models for planning, alignment, your alignment is shorter because you use AI to get all the information at once. Your code review is faster because you aligned up front and your coding is faster because AI did it. And so you're not actually really moving faster, but you're still reading everything, you're still reviewing the code.

Speaker 0: So, closing thoughts. Um, it's easy to hear all this and be a little burned out. Uh, you really like the world where we just let go of everything, and we can just like not have to ever read code ever again. Um, but we're engineers, and these are just constraints. Models are good at certain things and they're not good at other things. And so go figure out how to solve problems given a set of constraints. Use loops, they're great. Go solve hard problems. Seek leverage.

Speaker 0: Um, if you want help with this, we're building HumanLayer. HumanLayer is an AI IDE and collaboration platform, its building blocks for your software factory. Um, and soon to be a better verifier for software quality. We've got sort of a big move for cloud code and code style workspace. It walks you through the workflows for doing this sort of work. And we are talking to design partners. We are hiring founding engineers here and in San Francisco. Go and get these slides and go get them right now. You can try HumanLayer at humanlayer.com. Um, free for small teams. Go solve hard problems and complex code bases.

Speaker 3: Thank you all for your energy.

Speaker 4: Please welcome to the stage, the research scholar at Libaness Labs, Eric Admir.

Speaker 5: Um.

Speaker 1: Well, can you go back one slide? Sorry, all right. Good afternoon, everybody, thanks for being here after a long day of docs exhibits site event. Oh sorry, that was a site event. Um, I hope that um you have as much fun watching this demo as I had creating it, and let me first get this out of the way. This is not a product launch or announcement or anything. It's a twenty minute tutorial of how you can use elementary types, static types and compiler knowledge to make AI provably safe. And I'm sharing all my secrets with you today.

Speaker 1: And hopefully to have inspired some of you, next year you will have a booth downstairs where you have created a provably safe agentic harness. And or who knows maybe some of you have already solved it. Let me know, and then you know, we can grab a coffee instead of doing this talk. Um, with that out of the way, let's get going.

Speaker 1: Um, while I was prepping these slides, and I'm sorry that I was multitasking, but I was also um live coding on the side. And then when my attention went for a few seconds, because I was trying to convince the model to draw some pictures that it didn't want to do, and you will see some of these pictures later, you can guess which ones were rejected. Certainly, when I went to delete one of my files, and I'm sure this has happened to you before, um, or maybe not, maybe you always like, you know, run everything with no permissions. And then you say, "Yes, yes, yes," but I like to live dangerously.

Speaker 1: Um, but I'm convinced that if there's anything between the model's goal and where the model currently is, it will do everything that it can to reach that goal including killing us or deleting your files or deleting your database. So I think that these models are intrinsically very, very dangerous and we have to tame them. So that's what my demo is about.

Speaker 1: So let's like, you know, start this story. And it's I think a very, very sad story, but also a scary story of how we, as an industry, got to this point, where we are about to let normal people, the general public, give control of their computers, their finances, their whole personal lives over to AI agents. And we don't have any protection in place. I think very, very sad and scary.

Speaker 1: and so let me tell you the story how we got there. and i will like like have some characters like claude, and we will see dario and yeah as sam berne, but and the main character is is our friendly pit clot here. i think you can all remember um november thirty twenty twenty two. this was kind kind like a very special day in history, because this was the first time that you could speak to your computer. you could say you like summarize my emails, and it would, you know, answer you in perfect english. um i think for me, at least that was magic, but i think most of us didn't realize that by introducing this innocent looking function here, allow m that thinks a question and returns and answer that, that would open on door's box, and that would change our history forever.

Speaker 1: and but before we go continue the story, this conference is called ai engineer, right? so we are engineers and maybe we're the last generation of engineers that still understand what this is, what code is. or maybe most of you have already forgotten what code is, because all your code is written by agents. but if we look at this signature here, it says it the olympics question. tion, there's an answer. 对 the question and answers are not strings. they are very complicated json structures, and they get more complicated every day. every time a new rage of abs comes out. but for this talk, we can just assume that question and answer are just a big types. we we don't care about how care like we do care about what they represent.

Speaker 1: now anyanythe upheria of like these alalyams is being a great tools didn't last very long. and just when we thought that we have eradicated the small box of computer science equal ejection, it came back with vengeance because the bad guys discovered that you can trick alyams using protet injection and alyams have no distinction, make no distinction in coat and text. and so they are very, very easy to trick. and this, i think, is a bigger problem than sequal injection ever was, and but it was not prompt ejection. only that made alamam sckind of like have a bad rap. allaams are drained on the whole internet. and there's like a lot of good stuff on the internet, but also a lot of bad stuff. like how do you create a bomb? how how do sysyntisize ed drugs? how do you heck into lesystems? and the leaders of the big foundation labs, they got a little bit worried that that the government with interfere and regulated the industry. so they told there phd researchers go find a solution for this. this problem, right and quick it before solve it before, you know, the government steps um and here, the PHD types since their PHD types, they thought long and heart about the safety problem. and they came up with a new interface for elams. that's this kind of scary sheet on the right. look at that. well, does it say there's like some sick, my greek symbols, there's brobes ever ver. well, that is an probably you have heard of lean. anyone here heard of lean an is now the half thing right? like vc are are writing, like multibillion dollar checks. if you just say that you're doing something with lee, and of course, these beach d types researchers are using lean and you have to suffer because of that.

Speaker 1: now let's first look at the signature, and it's likely simply language called dephane. and what this thing says is that the alymtakes a question returns, an answer. it requires this question to be proper, which means that it's not an offended question. and then the model returns a safe answer. and this thing is proved automatically. so if you give it a proper question, it gives you a safe answer. now. i think there's too much attention for lean. i'm a recovering type holic and math addict, and i love lean, but there's many, many other dear provers and model checks out there like isaabi rock PVSDL plus and but lean is degrees that can't like keeps the vc money bumps going. so i will use lean and today. 嗯, so here years there kept luck it out the interface again and an and now and lean. you don't do automatically improving. if you're like a lean expert, you will say eric, well, we have grined and lean, but let's cup lucky up put that decide for a minute um. but in ine, you have to both show that how to compute the the result type and you have to do the proof by hand. so it's slightly different than the definitely a jump. but if you think about this thing for this a single nano second, you will realize that it's impossible to write a formal proof that an answer is safe or a question is proper. and that is why there are at least a hundred stargpps down here in exhibition hall that are using a allams as a judge, because this is not something that you can formally specifiabout. what does it mean that an answer is safe? that's not the mathematical property um. and of course, if you own an foundation model like these guys, you don't need external alalympse as a judge, you just be big it into the way. and you call it. the model is aligned. um but unfortunately, trying to bake alignment into the model is not foolproof, and models gets routinely jail broken, so they had to go to the pope. and ask it look get like, you know and bless their model that it's safe.

Speaker 1: i'm now i think it's terrible if, like a model said something offensive, but those are just words, and ultimately, words are like they just like, you know, they drip off your body. they don't do anything. some human has to act on words to make them dangerous. and so maybe that is what they mean by broadly safe. and when tropicc talk about safety and because it's still a human involved, but then something terrible happened, something really terrible happened that changed the world forever. and that is and june twenty twenty twenty three open. the eye announced two goal support in GPT four. and of course, all the other vendors rushed out to copy this. this is gold to the principle of minimum differentiation, and that is why all these aps look the same. now the act of adding two goals, ls ges ges AI safety from a philosophhical debate, do something that causes real danger. you could say two goals give the model gloss, in addition to a mouth, or you can say two gouses like having a gun alload a gun to them. but of course, nobody listens to me, everybody ignores what they say. and these guys just went ahead. and keft shift toocalls. they just yeah, this just do it. 我们在那个 now let's go back to like this is ai engineer and conference. so let's look at what is the difference in the signature of alalench when they add a tool calls. and it's just that little i owe there. and of course, it message up the formatof, the h nature. but if you look at the picture, there were you show, now certainly clolot goes from like a nice puppy to a dangerous thing. look at as all these dangerous tools. and now it's become scary, right? i've never seen anything scary er than an alalamma tool goals. now, if you look at this, this is like like a small step for a type whether giant leap for chaos, why is that? and that is because this io says that in order to compute the answer, the agent has to go to the agentic loop and is doing side effeture. while it's producing the answer, it might empty your bank account, but might delete your files. and then it give you a safe answer, but who cares about a safe answer, but all my files are gone, right? so that's right say, it's it's a giant leap for for chaos.

Speaker 1: um again, sorry. this is an engineer conference. let's look at this type io, and you don't have to understand it, but just see that there's a type theare called real world. yes, lean. this is a direct thing as a type called real world. and why is that because something of type io will mudate to the real world? so it warns you don't use this shit because it can make every versible side effects and like deleting your files, so solomon on hikes. and last year, disconference called an AI ent ent an allamb that's wrecking its environment in a loop. and i think he's a hero. i don't know if solo amon is here this year, but i think he should he deserves a round of a ploss because i think this is the right definition of an i agent. and by the way, this was one of the pictures that i had to robble to generate because it's clearly depics violence. and so it's kind of an unsafe, right? i have a picture that the big big violence. and so i we don't well, our agents have access to private data. they have interested content like the prompt ejections, and now we gave them tools. simon on versician caused this, the little try factor. and what can we do about this? well, um i don't know if you seen the dutch shocker fans, they have the famous march and days ic. do the left left left. do the left left?

Speaker 1: do the right right, right? this is actually the secret to solve this problem. the dutch team got eliminated yesterday. so you have to see me doing the dance, but all that we're doing is you 're pushing this io to the right to the right. and what you now see is that the tool belt of clolot goes to the left to the left, and certainly cllad is a nice puppy, right? because instead of executing the agenctic loop, it creates a plan and says, here is the plan to do the the genentic gloop. and now bernie will take that plan, and we will executed, and we all trust burniy. right? berney is a good guy. alright. so just do get to show it here. so in some that's what we're doing, we're air gapping the aggentic looloofrom the agent. so we don't let the agent run the agentic loop before the agent run it. we want to be able to jack it all right now.

Speaker 1: the problem is that if you get the value of type io of a, that's a really a black box, and the lean manual says, that is a black box, you cannot reason about it. so even though clad now gives us this plan, there's we cannot look into this plan. an doesn't allow us to do it. by the way. this is another picture, right? it. that's promotes drugs use and the model. let me do it. i'm a good teacker. yeah, i can just make it to forbidden pictures. so if we look in the lean, again, what you see here is that the model now compute an answer, but it doesn't compute the answer, right? it creates an io of answer. so this is a plan to generate the answer. and then it creates a proof that this that, that plan is safe. and the nice thing is here that you can get that, that proof without having to run the agentic c loop. but unfortunately, as i said, like this proof, if it's like something of type io, it's useless. i shit what can we do about that?

Speaker 1: so i keep going like moving you guys forward. and then we never get to the final answer, but there's one less trick. and you see the researchers here are becoming more much more sophisticated instead of the fatto do diones in the pnow there, like a real people. and what is better than than creating plplan of type io of a it's creating a program that's represents an expression of type io of a who that chounis very mata, right? i'm not matta in terms of meta AI. don't think there very meta, but and tta in the terms of like you like matter, you know, you know what i mean. and again, it's a small step for a signature, but a giant leap for safety. because now the model returns an expression, a program that resents your computation, if you know, linth or shshsharp well will recognize that this is one of the tricks that i always use. but if you know, list, this is, of course, second nature for you. and i cannot like, you know, have a talk without talking about monents. so if you ask yourself, what is this expression thing? well, that's just a moonat, but it's not just the monant. it's a free monant. where's a free monant? it's a monant that lost diydise.

Speaker 1: and now if you look at the the signature of property to proof that something and say, if you see that it takes an expression of a competition that returns an answer. and if you have taken any compiler course in college, you know that it's trivial to do data a low analysis, type checking and your own programs, right? so now we're safe. we are home safe and definitely wanted to remind you that we can solve the drive a problem just by doing int analysis on these expression on these programs. okay, this is the last gode that will show you because i'm running out of time. but just what to show you here that you know, you now have a simple inductive rigds of interpreter for this language, and you have a simple inductive proof, and the models can generate these proofs.

Speaker 1: 嗯, so too, um recapittoday. yeah, summarize, what we did is we went from unhinged allaams that was like, you know, could give bad answers to once that were aligned, then we saw how tools react ted, then we solve that by differing execution. so by air gapping, the the lamm from the tools and then the real solution was to refight the plan into a program. and the program that we could prove to be safe. now you would say, eric or 're a genius, no. and my brain is the size of a peanut. this is something that's called to proof carrying code. and it was invented by academics in the nineteen thousand ty. and and i'm just stealing it alright at the higher level, if you didn't understand the code, three point agents are dangerous and they'll proven safe. so if you never ever let your agents do something, unless you can absolutely prove that a safe and the language that this agents generated does not designed normal users, don't understand free monoment dodoes. it matter. it's a machine that consumes it. it's a machine that generates. it's a machine that proceits. so we should stop designing languages for humans, and it's all basic only requires programming one, no. and did we go? alright? that's it. um the end of the story.

Speaker 1: if you're curious to play with this, a bunch of academics and particular another, i'm in from harvard have implemented this. it's their own getop. it uses a slightly different language. but what i use it use is also a slightly different language than freeonage ts, then the ideas, the same the language. does it matters? it's the 嗯嗯 the principal that matter. so hopefully, you've learned tonight that it is actually possible to have mathematically proven safe agentic compute. and it only requires very elementary type systems and programming language. machininery. thank you so much.

Speaker 3: 啦啦啦啦啦啦啦啦.

Speaker 4: please welcome to the stage. the machine learning engineer model behavior at cursor lee robinson.

Speaker 0: 嗯, alright, everyone. i'm excited to be here excited to be back at ai engineer and talk a little bit about how our training models at curor. so how we trained the models and also how the model learns to train itself or a recursive model improvement. so our goal like cursor is to build the best possible ai models, which might make sense. you might have heard of this equation of if we just give the models more compute, we can get a better model out. and i think this is a helpful, simple vacation of the problem, but i want to actually click in a few layers deeper in the talk today and talk about all the different pieces that go into training these models.

Speaker 0: so we can think about this loop. we put a model up to the world, and then we get feedback from you all when you use the model, what goes well or places we can improve. we use that to scale and improve the data that we do for the next round of training. and we also then increase them out a compute and scale up training over all to make a new model. and this is a loop that can just go over and over again. however, if you see my helpeful sil or a turtle to bunny metre down in the bottom, right, it's pretty slow is is going to be a serial process. and you can only do and this incense one big run at the time. so we want to make this a little bit faster, but i'll actually go another layer deeper and add some more color here. there's actually two loops, the outer loop and interloop on the other loop, we have the feedback coming in, but we also have data like online metrics. so running ab tests and seeing what users prefer the feecheck point of a model that's going to then flow into, hopefully making better high quality evals that helping sure were getting the right behaviors we want out of the model. and also being able to create much more difficult problems for the models to try to solve where we then kind of shape the rewards that we want to get during training. so we want to climb that interloop as well.

Speaker 0: we have been training models for about a year at the large scale at cursor, and i want to talk about some of our progress so far. so we put out composer two point five in may, and it's now the most popular model and curser, which is exciting. and we scale up training here quite a bit by generating more are il environments, trying out some new methods for learning and also just making more ambitious problems for the models to solve. and the results have been pretty promising so far. like i mentioned, this is still a new effort for us. ml has really been in the blood of cursor since the start, where we were training more specialized models for things like tab or code hotto complete. but really, in the past year, we've sabfed up and build team with ambitions to train in know set of the art models, and we made some pretty good progress just in last two a months. 对对, people like composer right now. i think because it is both fast and pretty smart and also cost effective. and as we heerard from other speakers today, i think there is a space in the market right now for that type of model.

Speaker 0: in addition to also having the most intelligent models in the world, and we think is important to have a good selection of both of these type of things. So Composer, we think is serving a good niche here.

Speaker 2: Right, right, right.

Speaker 0: and when we released it, we were honestly pretty impressed with some of the public evals. It did a little better than we expected on Artificial Analysis. It was a pretty modest jump. However, there were a lot of behaviors that we found that we really wanted to improve the next version of the model. Notably, we wanted to have a much bigger and smarter model. We wanted to control every aspect of training. So ideally, doing a full pre-trained from scratch versus the previous open source space of Kimi that we were using. We wanted to infuse new data so that we can make the model great outside of more things in just coding, but more the general model. And then also just scale up every part of the training process, more data, more compute and really pushing our RL as far as we can.

Speaker 0: So first somewhat, I talked about improving the outer loop and then will drill in the inner loop. If you haven't used Cursor in a while, you might think about it as this IDE or tab autocomplete. And in reality, the vast vast majority of our revenue today comes from agent usage. And that means that all of the data inside of Cursor is also coming from agent usage. We use that to train better models. So for example, we kind of have two different buckets of feedback on the external side, when you're using the product, you can thumbs up or thumbs down different responses and get feedback. And we use that to then classify places where, for example, Composer's maybe doesn't do as good of a job. And we want to improve that for future versions. And then also on the internal side where heavy dogfooders of our models in our products, we're very critical and want to make sure we're using good models. And we, of course, use them all day. So we have a good mix of manual orts automated reports internally. And just lots of ways, we're trying to get the best behaviors out of the model. And if we do that over and over and over again, we can get better models out to the world.

Speaker 0: But really the place where we can make massive speed ups is improving that inner loop. So just to zoom back in and that again, we have these high quality evals. We have these very difficult training tasks, and we want to climb these evals as quickly as possible. So that we know if we make a new checkpoint of the model, we're actually making progress on the things that we want to measure.

Speaker 0: Right, right, right.

Speaker 0: So for example, some of the evals that we have introduced have already had things like understanding what you really meant when you have included maybe fifty scale files, it gets kind of hard for the models to figure out your actual intent or trying to figure out the line between when you push back and ask the user to clarify a question verses when you trust their judgment. And they said, no, I really wanted to do this. There's a lot of fine line and people have different preferences. So a lot of these evals are trying to shape a lot of those different behaviors and also model what it feels like to be a software engineer. We asked the models to do really ambitious things like, hey, we just had this eve. Could you have actually went and read through all the data og logs, read through Slack, read through Notion and came to the same conclusion or the same fix that we did. Um and a lot of models are just not very good at this that, and that um backs, a lot of evals that we create based on these soft engineering tasks.

Speaker 2: Right, right, right.

Speaker 0: Now, as the models get smarter, they also find very creative ways to hack the evals. So as we've been training for a new version of our model, we also notice there was some interesting reward hacking going on. The model learned how to just go back in the history and figure out if there was a solution or a part of a solution, they figured out good ways to go online. And if there was a public eval to see if there was a fork of the eval anywhere, they could look up the results from. And this affected our models as well as other models. So we did a little research here and found that if we did just a couple small changes on measuring public evals, we could have a pretty noticeable change in the scores that were reported. So first off, we would delete the git history at the start. We could restore at the end. So that would affect the run. And then also, we can have a network allowlist or just some basic controls on the sites that the agent can go and talk to. And I think this is helpful for public evals, which often are the things people are using to calibrate whether a model is good when it gets released. You know, you see that big chart of all the benchmark numbers, but this isn't really true tests of what it feels like to use these models. Like in reality, you have access to the internet, and that's why whatever you want in the internet with these models. So you're definitely using it. So you want to be able to test the true capabilities of the models, and that's why we have Cursor evals. You have a private eval set that is mostly made up of things that happen in a codebase, which is held out from the evals. So we ensure that the models aren't trained on it, and it's based on those real world engineering tasks.

Speaker 0: Now, another part of climbing that inner loop is trying to make very, very difficult problems for the models to solve. As the models get better, you might have noticed. If you're looking in an eval and all the models are scoring like ninety percent, probably time to retire that eval and try to get something more difficult. And the half-life of those evals will go down as models get smarter. And to do this, it requires a lot of things. It requires some amount of research taste, and what these problems should be. It requires a lot of compute. So you can try a lot of different ideas. Some of them are going to work. Some of them are not going to work, and there's a race against the clock here. So you want to try as many in parallel as you can. Just to put an example to this of one of those type of problems. Let's say that on the left, for example, you have each one of those dots is representing files in a code base. And one on the bottom you have the tests. One thing you can do is generate a very complex application or environment for a very ambitious application or task, and then you can delete part of it. You can delete a feature, you can delete files, and the test will then fail. And then you can ask these models to go and basically figure out however it wants to reimplement that feature, and it has a very verifiable goal of all the test passing to be able to get some reward back at the end. And this actually works out pretty well and has allowed us to scale making these interesting problems for the frontier models to solve.

Speaker 0: Additionally, we have found some new learning methods, which I personally think are really interesting. The first one is you can teach the model to kind of coach itself. So for example, if you think about an RL rollout or a conversation within an agent, this can be hundreds of thousands of tokens. And if you think about trying to grade at the end of this, where the model made a right decision or a wrong decision, it's kind of hard, right? You have all these tool calls, you thinking blocks. It's pretty hard to figure out where to assign that credit to the root issue. So the more precise we can be the better. You know, was it one of the tool calls or a thinking block? It's pretty hard. And one thing that we've done to improve this process is something called textual feedback. So we want to zoom in on one specific part of that rollout. And ideally we can hint or kind of nudge to the model. Hey, by the way, the way you could improve and then look at the probability again, and judge up the ones we want or downweight the ones we don't. Again, for example, on the left, you have this student case where you have a rollout, and it tries a tool call. And the tool call fails, it should have known that this tool was there, but it just decided not to work for this time. We can then use a teacher. We can use the same model, but we include this hint, and we say, hey, as a reminder, you have all of these tools available. And then we like I mentioned, we can just upvote or upweight the probabilities such that we can get the behaviors that we want. And this example is with the tool calling, but we can really use this for anything. We can use this for making style changes. We can use this to get any behavior. We want to influence the models during RL, and this is proven to be very valuable for us.

Speaker 0: Now how we scale these loops, both the inner and outer loops also comes down to scaling out compute. We announced back in March that we are partnering with CoreWeave to get access to a lot more compute and that allows us to train very large models from scratch, not only the product.

Speaker 0: but also the models down to the supercomputers.

Speaker 0: are the data centers where we're training these models with glosses and then increasingly to the chips as well with terifab.

Speaker 0: and that it just allows you to duce prepretty interesting things in taking advantage of that full stack.

Speaker 0: if you haven't seen glosses, i think it's really interesting. personally.

Speaker 0: they're able to train ah they are able to build out this supercomputer in one hundred and twenty two days for one thousand gp u, and then added another hundred hundred gp u in ninety two days.

Speaker 0: so very impressive. i had to do some pretty creative things to get this done and kind of take over this old factor.

Speaker 0: and in menmpice. and it's kind of shown they can stand up these data centors really quick, which is, of course, very for for our model training efforts.

Speaker 0: and for terarfab. i think it's also very interesting that they're building their own chips. i mean, to put the size of this into perspective. we just think about how large this physical structure is. i know you're all thinking it. it's like the size of a hundred buckies, which it's from my force from the south. you know, we love bucks ies, not even from the south. and i love buckies. this is like the crown drul of the south, the premium gas station as station. it's a lot of stuff, but that just puts it into perspective, the size.

Speaker 0: so going back to this equation at the start, more compute in, you get a better model out. i think it's sometimes hard to understand what does that computer even do? where do you actually put that compute? let's you have access to a bunch of mops, like what do i do with it? and i think it's helpful just to step through a few of the things, of course, first, you have actually serving the model to in users, but also you're serving up different checkpoints internally. you're running different AB tests. you're trying different variations of the model. you have the actual training process itself, but also the sub pieces from pretraining to bid training to rl. and then also, you're then training these derivative dels ls, do do parts of the proprocess clilimbing the interloop, which we'll talk about here in the second. you have the data generation. and the reward generation. so creating those really ambitious problems that i talked about, or when you're doing evalls trying to create these rurubricks for whether it was successful or not and give it some grade. and then actually judging those scores. um you also have the eballs themselves, ideally on every new checkpoint of the model you want to be continuously running evalls to see if you're improving in the places that you're measuring. and as well as just developing new evls ls all the time, like i mentioned, the the tough life with these emails as models to get smarter, you need to be really continuously investing in making these better. um and there's just the research itself. ideally, you want to free up your team of researchers to be able to tweet the nobs to try ambitious ideas experiment with new things as well as you side runs. and this all is compute that needs to be you allocated for somewhere.

Speaker 0: 嗯, but what that ultimately turns into is ideally, you can get in the state where you have multiple large training runs happening at the same time where the researchers are unblocked. and they can go try the research. and you're still kind of contributing back to this core fly wheel. and if you do that and we revisit our speemeter and bottom, right, you're starting to get to a point where you're getting something that's like. rs, i recuse ve model and improvement here where the models are improving much, much fast, sir. then the bottom neit becomes, how do you scale the folks actually train the models? how can you automate the more monotinous parts of machine learning or research so that you could get these useful models out into the world? and this is why i think it starts to get really interesting.

Speaker 0: 对, if you think about the model as mario, if you give it some tools, all the sudden you're more like a supermario. and if you give it great context, everything about your organization, all the places that you work, you connected to all your different tools, that context, kind of turns it into the firemario or the superfire mario? and just as kind of further prove this point and add a few examples here, i think, for tools, a lot of these are pretty obvious. the models can write code with with the harness. they can you run shocuman as they can look things up on the web. but i think increasingly, even with these primitive versions of memory, like writing files, these last three, i think you're just starting to become really popular and more useful, which is the models and the harnesses should be able to use a computer exactly like you would. it doesn't need m be be just inside of your gooe, or your c alii should be able to control every part of your computer.

Speaker 0: 对 you as a human on slack or on your tool is basically subscribing to slack threats in your head so that you can follow them for updates. ideally, you kind of want the models to just follow with read and then pink you, if it need ds. and just just like we have code bases that store the code increasingly, as these models do more work for us. and kind of need like a drop box for the ourselves, like, where do you store this? like decks, right? you could put that in code, i guess, but i think there's an increasingly new opportunity here.

Speaker 0: and and then for context, of course, you have all the different places you can hook up with mcps slack and nollinear data, dog said, uah and the code base itself. but i think he's last to are really interesting, which is increasingly we find that you have a human working with the team of agents, and then the agents can start working with the other agents. it's a little et a, but i think this will be a big trend in the next six months.

Speaker 0: just you kind of put an example to this. we've created these tools in these systems where researchers can run experiments directly from slack. we want to avoid this states of being bottenect on humans, launching and reviewing in babysitting runs. and we actually have an entire team just working on automating every part of the research work that isn't, you know, isn't freeing up the sesearchers time to work on the ambibious ous ideas. every perperson on ammteam gets access to this fleet of a, they can basically train models directly from slack. and what a few people on the team of taken this very far where they have these agent systems that can go and do a lot of work for them. maybe they want to go create a bunch of very difficult problems for the models to uh try and solve, or they want to create a whole bunch of new evalils based on some good ideas that they have, and they just want to let the models cook and go work for a while. but if something gets wrong, if the infrastructure goes thethere's some blip somewhere, the model can message them on slag or just page them directly and say, hey, this is really important. you don't want to lose six hours because your info rs down like you should go check this out right now. and this like human to agent coordination, i think, is just starting to be figured out, and it will be an increasing trend.

Speaker 0: the last bit here is that the model is learning to train the next model, and it's a little hard to wrap your brain around the way i like to think about it is every time you release a new version of this intelligence, then you can create or distill these derivtive versions that you use to speed up other parts of the training process, both the interloop in the other loop. so when you're trying to do your evels, for example, you have different models for doing the judging and you have your reward models as well. so when you make the top level model model smarter, it actually improves the whole system. if you think about the multiple training runs diagram, i showed you going to throw on a a new metre here, which is the brain to galaxy brain meter or the intelligence meader. you are a bottm eck here on the smartest model in your system. and if the smardest model then creathose dribit models, when you can improve that, you can actually make every single one of these loops much, much better. and you can raise the kind of floor the inteintelgence ence. and this is how you start to get to you improprolike this recursive self improvement this model that is just improving all the time on your behalf, and especially as we bring more and more compute online. i think this is really going to help us scale our model efforts and hopefully make even more useful models for you all to use to conclude.

Speaker 0: i just like to thank everyone on the cursor engineering and ml and research teams who have been working hard to get a new model out to you all here very soon, hopefully, very soon soon that we think will be pretty notable improvement over our last model, and we're excited for you to try it. thank you so much.

Speaker 3: 检察求你告诉你.

Speaker 4: ladies and gentlemen, please put your hands together and welcome back our mc member of the technical staff at keycard alii.

Speaker 5: how 啦啦啦啦？

Speaker 7: well, what it amazing first day of contact here at AI engineer.

Speaker 7: I think the biggest take away for me at today after listening to all amazing talks on software factories, is that the raw materials for software factories exist, right? Our context got larger, we have better memory model, improved vision model, we can actually verify the work. We've got better AI security best practices now to understand patterns to on agent identity and access control. But we have to have the discipline to be able to wield all of the raw materials correctly in order to build off our factories that actually work. And I really resonated with the nuances from the talks I heard this afternoon, such as a talk where I mentioned that one takeaway for me was that engineering as a practice is absolutely not dead. Engineers might not be writing all of the code anymore, agents might be writing the code, but engineering has to be part of that loop. Especially if you want to write software factories that actually work and don't produce slop. So I hope everyone learned a lot today. I know I certainly did, and thank you so much for your curiosity, your optimism, your energy. We will see you back here tomorrow.

Speaker 8: Thank you so much.

Speaker 9: This is our third year doing the Engineering World conference.

Speaker 5: We have so many events you are helping us step into the future of engineering. Engineering should look like forever, forever.

Speaker 8: Forever, forever.

Speaker 5: Yes, yes, yes, yes, forever yes.

Speaker 6: Let's see.