# You're Not Thinking Big Enough: Rebuilding Food Systems with AI Agents Cody Menefee, Firecrawl | AI Engineer World's Fair 2026 | 18:26 Source: https://www.youtube.com/watch?v=ztlXKfPCT3Q Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/youre-not-thinking-big-enough-rebuilding-food-systems-with-ai-agents Published: 2026-09-23 Tags: agents, human-in-the-loop ## TL;DR - Labor is the main reason more livestock is not raised on pasture, because rotational grazing requires daily decisions about fences, water, animals, and grass. - An AI system could combine GPS locations, drought data, grass height, and pasture history to suggest where a herd should move, with a farmer confirming the decision. - Open pasture knowledge, software that measures biomass and biodiversity, and open GPS collar APIs are needed before this system can work. ## Summary Cody Menefee argues that AI engineers should work on physical problems such as pasture-based livestock farming. He describes rotational grazing, where farmers divide land into paddocks and move animals every day so the grass can recover. GPS collars from companies such as Halter and Nofence can enforce virtual boundaries, but they do not decide where the herd should go. That decision still requires a farmer to inspect the grass. Menefee proposes combining animal locations, pasture history, drought conditions, grass height, biomass, and biodiversity data so an LLM can recommend the next paddock for human approval. He identifies three obstacles: collecting farming knowledge, building a vision system for pasture conditions, and getting open APIs from collar manufacturers. His Open Pasture project collects material from videos and research papers. He also describes stacking chickens behind grazing animals to manage parasites and improve pasture use. ## Key ideas ### Pasture-based livestock is held back by daily labor [04:41](https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=281s) Menefee starts from the assumption that more livestock should live on grass. He says 97% of cattle are finished on feedlots and 3% are raised on pasture, while arguing that more animals could be on grass. The main obstacle is labor. Rotational grazing requires dividing land into paddocks, giving animals roughly one day of feed, and moving them every day. Farmers must move fences, animals, and water while tracking what happened in each area. Leaving cattle to graze wherever they want damages pasture because they favor some plants, ignore others, and repeatedly trample the same sections. ### GPS collars control virtual fences but do not choose the next paddock [06:28](https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=388s) Halter and Nofence use GPS collars to create virtual boundaries and move animals remotely. Menefee calls this a useful way to extend available labor, but says the hard decision remains: where should the animals go next? Grass changes with drought, rainfall, and the amount of impact a paddock has received. Farmers currently walk onto the pasture, inspect the grass, and make an intuitive choice. They also need to keep grass in a productive growing stage. Grazing it too short slows recovery, while letting it grow too long makes it old and unappealing to animals. ### Different data sources trade off image quality, coverage, and practicality [07:50](https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=470s) Menefee describes three ways to observe pasture. Drone orthomosaic maps could provide high-resolution imagery, but farmers would need training and autonomous flights face regulatory barriers. He says no jurisdiction has approved autonomous drones for this use. Planet provides daily global satellite images at 1-by-1-meter resolution, but satellites are too distant to answer every question. A cheaper option is a trail camera pointed at a measuring object, such as a stick or apparatus, so an image can estimate grass height. These observations could provide the information that farmers now get by looking at the field themselves. ### An LLM could recommend grazing moves from several changing inputs [09:58](https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=598s) Menefee proposes putting an LLM in the grazing loop once pasture measurement and collar control are available. The model would consider animal GPS locations, where animals were the previous day, possible future movements, drought conditions, and grass height across the farm. It would reason about the best location for an individual cow, the herd, the pasture, and the wider farm system. The result would be a recommendation rather than an automatic claim of certainty. In his example, a farmer would review the suggested move before sending new virtual fence coordinates to the collars. ### The system needs a farming knowledge base and pasture vision [11:12](https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=672s) Menefee names three blockers. The first is a knowledge base containing practical information about when and why to move animals, which species work together, and how to build pasture biodiversity. He is collecting this material from farmer YouTube channels and research papers with Firecrawl, then placing it in his Open Pasture project. The second blocker is a visualization layer that can estimate biomass and track biodiversity over time. Biomass shows how much forage is available. Biodiversity matters because overgrazing can favor some grass types and reduce the range of nutrients available to cattle. ### Open collar hardware would let farmers compete on software [13:21](https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=801s) Menefee criticizes the closed model used by Halter and Nofence, where customers must buy the company's collars and cannot connect their own software. He understands the business reason, but wants an off-the-shelf collar with open APIs. His proposed system would let an LLM predict GPS locations and send those positions to the collar. He also points to farmers' practical attitude toward equipment and their frustration with locked-down agricultural technology. His direct request to the audience is to build a collar whose design and APIs are open, so independent software can improve the decision layer. ### Stacking chickens after grazing cattle can reduce parasite pressure [14:20](https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=860s) Menefee says the approach can extend beyond cows, sheep, and goats. He cites Pasturebird, which moves a chicken house across pasture on wheels every 24 hours. Chickens still need grain because they are omnivores, but their manure adds nitrogen to the soil. When cattle graze first and chickens follow, the chickens peck parasites from the cattle's droppings. Menefee says this can reduce the farm's parasite load and medication expense. Over time, he connects the practice with stronger breeding stock that needs fewer interventions. ### The useful AI problems are often uncertain recommendations, not deterministic answers [15:57](https://www.youtube.com/watch?v=ztlXKfPCT3Q&t=957s) Menefee closes by asking engineers to work on real-world physical systems that require many inputs. He says there is no single deterministic answer for the next paddock. There is a best guess based on current conditions, which a farmer can confirm or reject. He sees this human-approved recommendation pattern as a way to extend a farmer's labor and decision-making capacity. He also frames Firecrawl's work as gathering and packaging the context that agents need, while inviting the audience to contribute to Open Pasture or apply for jobs at Firecrawl. ## Notable quotes - "Pasture done right actually means moving animals constantly." (05:28) - "You have to know where to move the animals." (06:44) - "I need someone to make me a collar." (13:55) - "There isn't a next best paddock to move to. There's just your best guess on where you think they should go." (16:18) ## Tools & references mentioned - Firecrawl - Open Pasture - Halter - Nofence - Planet - Pasturebird - John Deere - Peter Thiel ## Who should watch - You build AI tools for software teams and want an example of a physical problem where the model must reason over changing environmental data. - You are working on agricultural technology, livestock management, computer vision, or GPS hardware and want to see where open software could fit. - You are interested in human-approved agent recommendations rather than fully autonomous decisions. ## Related talks - [Navigating AI's Frontier in 2025](https://aietalks.com/talks/navigating-ais-frontier-in-2025) (Grace Isford, Lux Capital, 17:55) - [Building Applications with AI Agents](https://aietalks.com/talks/building-applications-with-ai-agents) (Michael Albada, Microsoft, 15:50) - [The Age of the Agent](https://aietalks.com/talks/the-age-of-the-agent) (Flo Crivello, Lindy, 14:49) - [How agents will unlock the $500B promise of AI](https://aietalks.com/talks/how-agents-will-unlock-the-500b-promise-of-ai) (Donald Hruska, Retool, 16:22) - [How Anthropic Builds: Lessons from Labs](https://aietalks.com/talks/how-anthropic-builds-lessons-from-labs) (Mike Krieger, Anthropic, 26:11)