重构实体农业：利用多变量 AI 智能体破局草饲放牧系统

[Speaker 1]: Right? Hello everybody. My name is Cody. I put this picture up here because this jacket so far has not actually landed as well as I thought it would—no one gets the joke. So this was to really put it in front of your face. I don't just enjoy wearing heavily branded letterman jackets; we intentionally tried to lean into it a little bit.

So my name is Cody, I'm on the growth team at Firecrawl. And today, I'm here to tell you: you're not thinking big enough. But before I can actually get into that, I have to address a bit of an elephant in the room, which is Theo stole my talk. Theo put out a video about a month ago called "You Need to Think Bigger." But I would like to say I submitted the CFP for this talk long before Theo put his video out. I didn't steal his talk; he stole my talk.

Today, we're actually going to talk about farming. And yes, I actually mean farming. More specifically, I mean livestock farming, and even more specifically, I mean automating pasture rotation for grass-fed livestock systems. I'm feeling most of you did not expect to learn about cows and grass and farming today, but I'm here, so you're going to.

A little bit of background on who I am and why maybe you should listen to me: the short version is, I actually have no credentials that qualify me for this talk, but nonetheless, I want to do my best to give it. I grew up in Kentucky, have a background in blue-collar work—I was a bartender, a mechanic, a server, a whole bunch of things. Never actually a farmer, though. And then I sort of found my way into engineering and software development. I actually don't really like taking the title of software engineer; I'm pretty averse to that, it feels like stolen valor, because I am the vibe coder most of you all are scared of. I use AI agents all day long, I don't have syntax memorized, I'm not proficient in any particular coding language, but I will crank out some stuff on a weekend. Today, I work at Firecrawl. We're building context for AI agents. We have a series of web data APIs to give your agents access to the web. Again, I sit on the growth team.

But today, we're actually going to get into some more farming stuff. First, a bit of credential I do have: this is a real picture of me hauling turkeys on top of my Tesla. And I do still have a crack in that glass ceiling because of it. This was in Nashville, where I live in the middle of a residential neighborhood where you are not allowed to raise turkeys. But I raised ten turkeys in my backyard because I wanted to know what it was like to actually raise livestock myself that I would eat. It's a very mentally difficult process, to be honest. But this was me loading them up on the roof of my Tesla, and then I drove for three hours with them to the processor, had to stop at a Supercharger on the way, and lots of people were taking pictures. What I can tell you is your range is sufficiently decreased when you have a giant windbreak full of turkeys on top of the roof. So that was quite the anxious drive.

I can tell you, I also almost left engineering to be a farmer. I really wanted to raise chickens. This is a real product image that I came up with: I wanted to wrap turkeys in white wrapping and literally just slap "TURKEY" on top of it. I thought it was provocative; I thought it would get you to buy chickens. But I realized it's actually really hard to make any money farming—surprise, surprise. And I have a wife, I have two kids. It didn't feel right to ask them to give up the lives they had so that I could go cosplay as a farmer and raise chickens. So I decided to pivot and see if there were ways that we could scale farming itself, and that's the type of systems that I'm interested in when it comes to livestock agriculture.

So three things I want to accomplish in this talk:

1. Convince you all to pursue bigger ideas. I think at a lot of these conferences, a lot of us individually spend a lot of time talking about building software for people who build software for people who build software, and so on and so forth. I'm really just here to challenge you that there are other problems to solve than just another MCP for another SaaS solution at another company.
2. I'm just trying to take advantage of a captive audience. If you corner me anywhere at any time, there's a good chance I will talk to you about farming. So here I am.
3. Hopefully, I can convince you to come work with Firecrawl.

[Speaker 1]: First things first: I believe livestock belong on pasture. I think animals should live on grass. I think it's better for the animal, the consumer, the farmer, the ecosystem. I could give you a whole TED Talk on each of those if I need to; you can find me later if you need me to tell you why it's better for animals to be on grass, but I don't have enough time to get into all of that. Take my word for it. Let's start there: the assumption is animals should be on grass.

This is the ball I want to hit. I am not actually anti-containment farming; I think there's a reason we needed to do that. But 97% of cows are still currently finished on feedlots; 3% are raised on pasture. My contention here is more animals could be on grass. I want to try to figure out how to get animals on grass.

The question is, why aren't they on grass? And that is: labor is the bottleneck. It is a pain in the ass actually raising animals on grass. Pasture done right actually means moving animals constantly, and that takes a lot of work.

If you think about grass-fed beef, you might think of: I have 100 cows, 100 acres; I put 100 cows on 100 acres, they ate the grass, got beef at the end of the year. That's not quite how it works. You will very rapidly decrease the quality of your pasture if you just let cows graze where they want, because they'll graze their favorite things, ignore things that they shouldn't, trample areas consistently, and so on and so forth.

So the solution to that is rotational grazing. What this means is you break up your pasture into individual paddocks where the animals have enough food for one day, and then you move them every single day. This allows certain areas to rest and other areas to be grazed, and over time will increase the efficacy of your pasture. But this takes a whole, whole lot of work. This means you have to move fences, animals, water, and keep track of it every single day in order to appropriately move the animals as often as they need to.

There are some solutions actually trying to work on this problem. You may have seen a company called Halter in the news recently; Peter Thiel invested at a $2 billion valuation. NoFence is the other company. What these companies do is provide collars for the animals connected to GPS satellites, allowing you to draw virtual boundaries where you can move the animals remotely. I think this is a great step in the direction of trying to expand labor.

But this has a problem, which is you have to know where to move the animals. This is not a science, to actually be honest with you. You can't move them in a straight line across the pasture routinely every single day to the same part of land. The reason is that grass doesn't grow the same every single day. There are drought conditions, rainfall, how much impact a particular section of the paddock has had. And the way this is solved today is actually farmers going out on pasture, putting eyes on the grass, and making intuitive decisions about where the next best move should be.

So the question is: how do we replace the farmer's eyes on pasture so that they can remotely make educated decisions on where to move their virtual fences? There's a bit more that actually goes into this as well, and that is you have to also know how tall the grass is. Grass has a growing cycle: if you graze it way too short, it takes a long time to come back; if you let it go too long, it becomes old and bitter, and the animals don't like it. There's this juvenile sweet spot that you want to keep the grass in. You want to cut it before it gets too tall, but you also don't want to cut it too short. You need to keep the animals moving and then constantly coming back to the same pasture so that your grass stays at the most optimal growing phase and constantly has the most productivity possible.

[Speaker 1]: So a couple of ways that we can do this—these are things, these are my solutions, something I've actually been working on and thinking about how we can do this.

A couple of options that I have are:

1. Drone orthomosaic maps. If we could find a way to automate drone flights where we go fly around our pasture, take a whole bunch of pictures, get some very high-fidelity, high-resolution images of the grass that farmers could analyze. The problem with this is there's a lot of skill upgrading we need to do with the farmers to teach them how to fly drones, a lot of regulatory issues for keeping the drones in sight, and ideally this would be autonomous. Currently, there isn't a jurisdiction in the world that has approved autonomous drones for these types of applications. So this is a really big bottleneck. I think it has pretty high fidelity and quality of imagery, but it is going to be a hard problem to solve in terms of actually getting all those hurdles accomplished.
2. Satellites is my most favorite option today. There's a local company called Planet out there taking pictures of the entire globe every single day with 1x1 meter resolution. But they're still satellites really high up in the sky, and it is hard to tell some of the things you need to tell to actually make those educated decisions.
3. The middle photo here is actually from a friend of mine out in Missouri working as a research graduate assistant at Lincoln University. This idea is just putting a trail cam next to a tree and some measuring apparatus that that camera can look at, and just figuring out: how tall is the grass in relation to that particular object? Just so that we have some sort of reference point that we can use to see how well the grass is growing back.

If we can solve this problem along with the collar situation, I think there's a world here where we can drop an LLM in the middle of this loop and start to work on autonomous grazing operations. What this would mean is an LLM essentially making the next best decision on where the animals should be on any given day. This is multivariate data analysis; this requires the LLM to have several data streams: where animals are, GPS locations, where they were yesterday, where they might go tomorrow, what the drought conditions are in the area, how tall the grass is across the entire farm.

And actually, it has to make this decision not just on a day-to-day basis, but in varying degrees of relation:

• Where is the best next place for a particular cow to be?
• Where is the best next place for the herd to be in relationship to the pasture itself, in relationship to the farm as a whole?
• And then more broadly, the ecosystem at large.

All of these components feedback into each other. If you can optimize this entire picture, you have a more productive farm. We can actually have more animals on fewer acres, which is how we end up actually scaling to compete with the feedlot style: we can actually have more cows on fewer grass.

[Speaker 1]: How do we solve this problem? There's a couple of components. There's three main blockers that I think exist in order for us to actually create this system:

The first one is building a knowledge base, and this is primarily what I'm working on at Firecrawl, and then an open-source project I have called OpenPasture. The idea here is a lot of the knowledge on when to move, why to move, how to move, the benefits of moving, etc., is all locked up primarily in videos. There's a bunch of really cool farmers out there—I could give you a whole bunch of channels you could go down rabbit holes on—just good old guys out in Missouri, Tennessee, Kentucky, trying to move their animals every single day, telling you what they're learning, telling you what species are best for this, what legumes you want to aim for in the wild diversity of your pasture, and a whole bunch of things that go into this. We need to build that knowledge base. Firecrawl is a toolkit that I use to actually collect this data, going out scraping those YouTube videos, scraping research papers such as arXiv, building this knowledge base up. OpenPasture is the actual repository I put this information in to make it available to any farmer that I think might be able to use it.

The next thing to solve is the visualization layer. There's a lot of components we need to know about the grass that farmers primarily get out of intuition from looking at the pasture. The two main things worth figuring out about the pasture—both where the animals are, where they should go, where you want them to be—is:

1. What is the biomass? How much foliage actually is available for them to consume?
2. In the long term, what is the biodiversity of that particular pasture? If they overgraze sections too heavily, they'll start to over-index on different types of species, and warm-season grasses and legumes degrade, etc. Ideally, you want a really rounded, really diverse pasture over time to make sure that the cattle are getting the nutrients they need to, so you don't have to supplement with things like copper, minerals, etc. Ideally, they get all of the micronutrients and macronutrients from the grass itself, which becomes an entirely, ideally, hands-off system.

And then the third one is those geofence companies—NoFence, Halter. While I appreciate the technology they're trying to push forward, I have a pretty strong disagreement with them, which is: in order to use their software, they require you to buy their collars, and you can't plug your own software into their collars. From a business standpoint, I get why this is; from an industry standpoint, I think it's a real pain in the ass. I would like to innovate on the software layer. I would like to push GPS locations to these collars that my LLM can predict. I don't want to have to rely on their software to do this because I don't think it's as good, or I think I can make it better, to be totally honest with you.

So a bit of the purpose of this talk is actually a call to action for you all in the audience: I need someone to make me a collar. I need it to be open, the APIs need to be open. Ideally, it's an off-the-shelf solution, some component parts that we can slap together. Farmers are pretty scrappy and like to build their own things. There's a lot of analogy to John Deere and the Right to Repair. So I ask anyone looking to design an open collar where the patent can be open and the APIs are open so that we can compete on software and optimize this solution.

The next thing I'd like to tease you about is: this actually goes beyond just ruminants. Ruminants are a type of animal—cows, sheep, goats—all ruminant animals that chew grass and digest it in an organ called the rumen. But there's actually an additional benefit we get when we stack species on these pasture rotations. A company called Pasturebird was a big catalyst for me to really get obsessed with this idea. What they did is took your normal chicken house, put it up on big wheels, and automated the movement so it creeps every 24 hours across pasture. The reason you can do it this scientifically with chickens is because they don't actually get most of their nutrients from the grass—you have to supplement them with grain feed because chickens are omnivores, not quite herbivores. So you can just inch this coop across the grass, giving them fertile land to grow on. The nitrogen from their droppings actually helps as a fertilizer for the grass itself.

And there's an added benefit here when you stack the ruminants with the chickens: if you run ruminants first, they sort of cut off the top of the grass, and then chickens come behind them and peck out the parasites from their droppings. It reduces the parasite load overall across your farm, which reduces the medication expense you have to pay to keep your animals healthy. Over time, you have a more robust breeding stock, so you have stronger animals over time that require fewer interventions and can be left alone to just eat grass and turn into meat eventually.

[Speaker 1]: So the three things I really hope you will take away from this talk:

1. I think we need to find big, real-world, physical problems that we can solve that require multivariate analysis and not quite yes/no decisions. There are lots of problems out there that don't actually have deterministic solutions. I hear a lot of engineers talk about how we turn LLMs into deterministic processes, and my contention is actually there's a lot of problems that you can't solve with deterministic algorithms. This is one of them. It's a multivariate analysis. There isn't a single "next best paddock" to move to; there's just a best guess on where you think they should go. I think we can take systems like that, that are multi-data input systems, and drop an LLM in the center that actually reasons over the data and at least makes a suggestion that the human can confirm or deny. We can really start to scale systems like this that are very much restricted by the farmer's ability to allocate their own labor or their decision-making power, and really give them the tools that they need to grow their operations. Hopefully, I think all animals can be raised on grass if we solve these problems.
2. The last one is: maybe you can help me solve some of these problems. The number one problem is actually just giving the agents the context they need—gathering the data, packaging that data, and then presenting it in a way that the LLM can reason over. That's what we do over at Firecrawl.

So you're not thinking big enough. Firecrawl is building a context layer for AI, and I hope you can come build with us. We're hiring, so here's all the job postings we currently have. Go on our website and if you find something that works out for you, reach out to me. We'd love to have more people trying to figure out how we get data off the web to solve some of these more complex problems and present that data as context to the AI agents. Perhaps we can make the world a better place.

My name is Cody. OpenPasture is my open-source project; Firecrawl is where I do my day-to-day life. And these are my socials. I'll hang around for a little bit. I'd love to chat more about animals, cows, birds, all you like. Thank you very much for coming.