Tjisk Kumar: Good morning. Thank you. Thank you. Good morning. We are here. Look at this. It's a full room. We are so excited. How many of you enjoyed yesterday? My highlight was Malta yesterday when he shared this slide of how Europe is leading AI innovation. And I think this is so cool because oftentimes we feel like the underdog, and this was so validating that real innovation is coming out of Europe with even the DeepMind office in Berlin.

Tjisk Kumar: Today is day two and we're going to talk about a lot of interesting topics. So, we're going to talk about coding agents. We're going to talk about MCP. Who's using MCP here? Look at that. Almost everybody. Incredible. We're going to talk about AI architecture, media, GPUs. But before we do that, will you join me in giving it up for our sponsors, Google DeepMind, our presenting sponsor? Let's also give a huge round of applause for our platinum sponsors. We got Brain Trust, WorkOS, OpenAI.

Omar Santoro: All right. Hi everyone. I'm super excited to give this talk because just seven days ago we released Gemma 4. So GM is Google minds a family of open models. Open models means that these are models that you can take, you can download, you can run in your own infrastructure, your own devices, you can fine-tune for your own use cases.

Omar Santoro: So, Gemma 4 is the family of most capable of open models that Google has released ever. These are models that go from two billion parameters all the way to 32 billion parameters. The smallest two models can run in an Android phone, in an iOS, in a iPhone phone as well, even in a Raspberry Pi. These are really small models that are multimodel have reasoning can do like very cool ondevice agentic things. Then you have the 31B that's the most intelligent model the most capable. But even the 31B is a model that can run in a consumer GPU.

Omar Santoro: So actually Yema E2B has a new novel kind of architecture called per layer embeddings. The TLDR here is that pretty much there is like a embedding kind of per each layer, and it works more of a lookup table rather than a computation that you need to do. So pretty much this is a extremely fast thing. You don't need to have this in the GPU. You can have this in the CPU. You can have this in the disk. leveraging lama CPP with a simple flag override tensor and then you move the per layer embeddings to CPU or even to disk. With the M4 we change our license to an actual Apache 2 license.

David Sora Par: Okay. Well, welcome. This is an MCP application. That's an agent shipping its own interface, not through like a plug-in, not through an SDK, not rendered on the fly by the model on the client side or hardcoded into the product. That is something that is served over an MCP server. And you can take the server, put it into cloud, you can put it into chatbt, you can put it into VS Code, cursor, and it will just work.

David Sora Par: For doing that you need semantics. You need to have both sides the client and the server to understand what each side is talking to understand how you render this. And for that you need a protocol. We're now like at 110 million monthly downloads. Just a bit for context, React, one of the most successful open-source projects probably of the last decades, took roughly double the amount of time to reach that download volume.

David Sora Par: But I still think this is just the absolute beginning because I think 2025 was all about exploring in 2026 is all about putting these agents into production. We're going to have general agents that will do real knowledge worker stuff like things a financial analyst want to do. And they need one thing in particular: connectivity. In my mind, there are three major things that you want to consider building an agent in 2026. It's skills, MCP, and of course, like CLI or computer use.

David Sora Par: The number one thing we're going to do there is that we need to go and start building something called progressive discovery. Everybody so far has done is to simply put all the tools into the context window. But what you should do instead you should start using this progressive discovery pattern which is to say use something like tool search to defer the loading of the tools. We have a proposal from our friends at Google who are working with something called a stateless transport protocol which make it significantly easier to just treat MCP servers like another stateless rest server.

Ido Salamon: So good morning London. agents are amazing. Uh, but if one agent is so amazing, why don't we scale up to 10 or 20 or 100 different agents and be a 100 times more amazing? It is pretty simple. spinning them up isn't the problem. It's us. We are the bottleneck in orchestrating all of these agents.

Ido Salamon: Luckily, they're not really brand new. If you're a gamer or used to play games, managing dozens of units probably sounds a little bit familiar, which is why I built AgentCraft, which is an orchestrator that aims to raise the ceiling of human agent collaboration by taking learnings from gaming and transferring them into productivity.

Ido Salamon: this is a physical manifestation of a coding agent. if we look at the map, you would notice that it's actually a projection of my file system. I can actually track and see visually what the agent is working on, which file. why not just create a heat map? I can actually try and see visualize collisions and I can even prevent them proactively. We can simply use muscle memory to quickly cycle between the agents that need our help.

Mario Zechn: Hey there, I'm Mario. I built Pi in a world of slop. In the beginning there was cloud code and was good. Back then, it was simple and predictable. But eventually the team got bigger and they built a lot of features. With more features come more bugs. The real problem is that my context wasn't my context. Cloud code is the thing that controls my context. And behind my back, cloud code does things to the context.

Mario Zechn: So that's Pi. It's an agent that adapts to your workflow instead of the other way around. It comes with four tools. Read, edit, bash. Here's Pi's system prompt. That's it. The models are actually reinforcement trained up the wazoo. So they know what a coding agent is. You don't need 10,000 tokens to tell them you're a coding agent.

Mario Zechn: OSS in the age of clankers. Clankers are destroying OSS. The problem is that agents are emergent learned complexity. 90% of code on the internet is our old garbage. And that's what the models learn from. Agents will happily keep [ __ ] into your codebase. So the agent patches locally and [ __ ] up globally. You cannot trust your codebase anymore and also not your test because your agent wrote your test.

Mario Zechn: My final slide, slow the [ __ ] down. Think about what you're building and why. And don't just build because your agent can do it. Learn to say no. Critical code, read every [ __ ] line. Write it by hand. And that friction is the thing that builds the understanding of the system in your head.

Armen Ronacher: Today I want to talk with Christina about friction a little bit. I have created Flask, which is a Python framework, which ironically is so much in the weights that a lot of people are learning about it now because the machines are producing it. We have been building with or on agents for a good 12 months. We had huge leverage and great disappointment.

Christina Ponella Cubro: I am what I like to call a native AI engineer. what this means is like they've been super foundational in how I've become a software engineer. these tools are super addictive. You never know if that next prompt is going to be the one that makes your product work or if it's going to be that last drop of slop. because we produce a lot of output very fast, we are tricked into thinking that we're actually being more efficient.

Armen Ronacher: the agents are writing kind of code that when you as a human start learning how to write code you wouldn't necessarily write. they optimize towards making progress towards shipping stuff. And as a result, they're creating many more failure conditions than human written code normally would do. because you as a human feel a little bit bad when you write code like this. But the agent doesn't have a reason for this.

Christina Ponella Cubro: what we're proposing here is an agent legible codebase. one of the main points is modularization. we enforce unique function names. And the reason for this is not just more legibility, but it's actually also the token efficiency. we've started exploring something recently called erasable syntax only TypeScript mode. there's one source of truth between your actual code and the compiler.

David Gomez: I'm going to be talking about how markdown is basically the new code. we recently replaced a lot of code in the cursor application with just markdown just a skill. Like I think it was around 15,000 lines of code deleted.

David Gomez: We decided that there are two primitives that we could use to effectively allow cursor users to use word trees. One is Asian skills and the other are subentations.

David Gomez: The skill is written is actually really simple. it's basically a set of instructions telling the model how to create word trees. We have to instruct the model to stay on that work tree. we do that with some aggressive prompting effectively. it's effectively a way for you to compete on have have different models compete on the same task. And then you can even preview the changes.

Sarah Chang: over the past few years, we as developers have developed a series of bad habits as a result of slow AI code generation. Codex Spark can generate code at 1,200 tokens per second. So in this new era, this is 20 times faster.

Sarah Chang: at 1200 tokens per second, a model like codec spark makes validation basically free. There is no excuse why you should not be doing things like test suites, linting, diff reviews. Another tip is exploring cherrypicking. I can generate five sub aents that are each generating 15 versions and now I have 75 versions and I pick the one that's best.

Sarah Chang: You should view it much more as a pure programmer. sit down with it and actually steer it. You can do things like ban the model from deleting files, give it a max diff size. a good way that you can think about how do I externalize this memory is with this four file system: agents MD, plan MD, progress MD and verify MD.

Lawrence Jones: our goal is actually to fully automate production investigations. But behind this investigation is like hundreds, if not thousands of prompts. So, how on earth do we scalably understand how this system is performing? They are now complicated enough that you can't as a human really tractably dig into how these things are performing.

Lawrence Jones: can we just download all of the UI that we have as a file system? And that's kind of what we've done. these agents are fantastic at using file systems and just going through this data. You sit there in the session and you go hey like have a look at this. Tell me what you think has gone wrong. and then it will be able to tell you where you should be making the modification.

Lawrence Jones: we created this repo called scrapbook. we parallelize out all of your agents. they can all individually build their analysis of an investigation. and then you go into the next stage of the pipeline where you do some cohort clustering. by clustering it together, you end up with actually a really really useful report.

Luke Alvo: the bottleneck in software engineering nowadays is not intelligence. It's now limited by human attention. we kept asking ourselves, what if a human decides what to build and then a system figures out how to do so? An agent could just work for hours for days and you come back to finish work.

Luke Alvo: notably, a mission is not a single agent session. It's an ecosystem of agents. It uses a three-role architecture: orchestrator, workers, and then there's validators. the orchestrator produces a plan that includes features, milestones, and then something that's called a validation contract. workers handle implementation. and then the last role are validators. They handle verification.

Luke Alvo: our longest mission ran for 16 days. The difference with missions is that we run features serially. Within a feature, we allow for parallelization on readonly operations. It seems slower on paper, but the error rate drops dramatically and when you have tasks that run for many days, the sort of correctness compounds.

Gergely Orosz: things are trending the wrong way right now. what happens when when agents are capable of doing everything immediately for you? the pendulum has swung too far into the wrong direction. Uber was a winner takes it all market. you just had to ship immensely. what I saw at Uber was that hyperrowth that I never want to go through again.

Thomas Artman: great products come out of saying no to you know 999 things and yes to one thing and with AI we might be in a place where it's just too easy to say yes. We still say no to a lot of custom requests. 10% of our bugs are automatically fixed by a singleshot AI instance. I do foresee a future where like it gets closer to 100%.

Thomas Artman: every engineer would show one fix that they did was quality and it went from like a one pixel change to oh I just made our backend way more efficient. people kept on finding problems in the product because they knew they had to come to the next Wednesday meeting with a fix.

Thomas Artman: zero bug policy literally means that if a bug gets reported it gets assigned to somebody automatically and that becomes your highest priority. the rate at which you have to fix bugs is again constant. It doesn't matter whether you fix them two months from now or immediately.

Jacob Laurson: planning work and reviewing work is the new bottleneck. doing the actual work is extremely cheap. if you've ever worked with a longunning complex agent, you get a new contract. Does it look was it only clause three that was changed? Probably not. And so you end up in this state.

Jacob Laurson: verifiers rule states that if it's a task is solvable and it's easy to verify then it's going to get solved by AI. verticals have tasks that are different places on the spectrum. Writing a contract is very easy to solve, but actually extremely difficult to verify. You can decompose task. leave picking risk profile to the human, but I can try to get other stuff down where it's easy to verify.

Jacob Laurson: Chat is one-dimensional. It's a very low bandwidth interface. So what's a better interface? Well, I think humans and agents should collaborate in high bandwidth artifacts. durable interface where it makes sense to collaborate. agents aren't humans and so we should not constrain them to human language.

Peter Gstiff: any benchmark you seem to look at line goes up. this could create this kind of psychosis that we all see. I think there's still quite a few things missing. the idea behind the [ __ ] benchmark is quite simple. what happens if you ask nonsense questions?

Peter Gstiff: quite often it would maybe have one line where it would question the premise of this question and then spend 20 paragraphs trying to solve it. the reason why that happens is that they were trained so much to solve the task at any cost. 9% of the time people would get two responses from two good models and they don't like them. whenever you try to build games with LLMs, it just feels like they have no idea like how to build actual games.

Swix: I start introducing a workflow of our contract designer showing me a Figma page and asking me to turn it into reality. I just added Dev into it. hooked up Devon to Figma and in very short order we have a perfectly functioning website that is pixel perfect to the Figma.

Swix: I am getting more work out of my employees because they enjoy doing it because the feedback cycle for them from waiting blocking on me is gone. this entire schedule is managed by Devon. whenever someone comes in with a speaker change, I just say, "Devon, handle it for me." I asked Devon to research where can I get a lobster in London. that's literally the the lobster that you had was bought from Devon.

Swix: once you get enough psychosis, you are thinking about replacing entire pieces of SAS. here is me arguing with my employees about kicking out a SAS tool and building it ourselves because we can. 60% of the user base of Versell now is bots is agents. It's not humans. So actually your dashboards don't matter. Your APIs matter your CLI matter. Your MCPs matter. Agents for everything else are coming.

Tjisk Kumar: We did it. Y'all are such an amazing crowd. What a journey. Give it up for your speakers each and everyone. we have a party coming up at 7:00 local time in a club called Fabric. But here's the deal. It's not clubbing. we're going to create an atmosphere that's not you can talk to each other and ideally you do. come with your badge. let's move to taking a group photo together.