You’re Not Thinking Big Enough: Rebuilding Food Systems with AI Agents — Cody Menefee, Firecrawl

Cody Menefee, on Firecrawl’s growth team, opens with a jacket joke and a mock dispute over who first claimed the injunction to think bigger. His actual subject is unusually specific for an AI talk: automating pasture rotation for grass-fed livestock. The motivation comes from trying to raise animals himself, then discovering how difficult it would be to turn that interest into a living.

The memorable credential is ten turkeys raised in his Nashville backyard, in a residential neighborhood where he says that was not allowed. Raising livestock he would eventually eat was mentally difficult. Getting the birds processed added a more literal engineering problem: he loaded them onto the roof of his Tesla and drove three hours to the processor. The load damaged the glass roof, and a “giant windbreak full of turkeys” reduced the car’s range enough to make the journey anxious, with a Supercharger stop along the way.

He nearly left software development to raise chickens. Farming’s economics, and the consequences for his wife and two children, changed that plan. Rather than ask his family to support what he calls going to “cosplay as a farmer,” he began looking for ways to scale the livestock systems he cared about. That pivot gives the talk its challenge: an industry busy building software for people who build software could also tackle the work that keeps physical businesses small.

The starting judgment is Menefee’s: livestock belongs on pasture, which he believes benefits animals, consumers, farmers and ecosystems. He leaves that case largely undeveloped and acknowledges a role for containment farming. He cites 97% of cattle finishing on feedlots and 3% being raised on pasture; those figures have no stated geography or supporting study here. His practical question is how to put more animals on grass, with labor as the bottleneck he wants to address.

Consider his example of 100 cows on 100 acres. Simply releasing the cows and waiting for beef at the end of the year gives the animals control over where grazing happens. They repeatedly eat favorite plants, leave others alone and trample the same areas. In his account, the observable change is declining pasture quality: access to plenty of land does not ensure that grazing pressure is distributed in a way that lets the land recover.

Rotational grazing changes that sequence. Divide the pasture into paddocks with enough food for one day, let the herd graze one paddock, then move it the next day. The area just grazed rests while another area supplies feed. The desired improvement comes from controlling both grazing and recovery, rather than letting the herd revisit whichever plants it prefers. But the daily move includes fences, animals, water and recordkeeping. A better grazing pattern creates a recurring operational burden.

GPS collars offer a way to change grazing boundaries remotely. Menefee names Halter and Nofence as examples of companies supplying collars that let farmers draw virtual boundaries and move animals without relocating a physical fence. This removes some work from the daily rotation, but it leaves a decision upstream: where should the new boundary go?

The 100-acre example now needs more than a fixed route through its paddocks. Grass does not grow at the same rate everywhere or every day. Drought, rainfall and the impact of previous grazing change which area is ready. A farmer currently goes onto the pasture, looks at the grass and uses that observation to choose the next move. Remote fence control becomes much more useful if the farmer can also inspect the pasture remotely.

Grass height also affects when the herd should leave and return. Grazing too short slows recovery; leaving grass too long makes it old and less appealing to the animals in Menefee’s description. The target is a “juvenile sweet spot”: graze before the grass becomes too mature, leave enough for it to recover, then return when it is ready. Yesterday’s move therefore changes tomorrow’s choices. The system needs observations over time, not just a map of available land.

The observation options trade image detail against the work needed to collect it. Each offers a different answer to the same question: what can the farmer learn about grass without going out to look?

• Drone orthomosaic maps: Fly over the pasture and collect many photographs to produce detailed imagery. The attraction is high resolution. The burden is teaching farmers to operate drones and meeting flight restrictions, including keeping them in sight. Menefee describes autonomous operation as the desired end state and regulatory approval as a major obstacle; his worldwide claim about approval is not established here.
• Satellite imagery: Repeated overhead images could avoid the farm-level flight operation. Menefee favors this option and names Planet, but distance limits the detail available for grazing decisions. The daily global coverage and one-by-one-meter resolution he cites are unverified specifications in this recording.
• Trail cameras with a measuring reference: Place a camera beside a tree and a measuring apparatus. Compare the grass against the visible reference to estimate its height and follow regrowth. This is the simple local option: the image has something against which growth can be judged.

Once pasture observations and collar control are available, an LLM could combine them to recommend the next grazing location. Its inputs would include the animals’ GPS positions, where they were yesterday, possible future destinations, local drought conditions and grass height across the farm. The proposed system has unresolved components rather than a validated autonomous deployment; increasing the number of animals supported on fewer acres is its goal, not a demonstrated result.

The decision also spans different scales. A good next location for one cow must fit the herd’s needs, the pasture’s recovery and the farm’s longer-term condition. Returning too soon may provide feed today while weakening future regrowth. In the developing pasture example, the useful change is from following a routine route to choosing a move informed by current grass and prior use. Menefee’s closing version makes this a suggestion that a human can confirm or deny.

What has to connect before a recommendation can change grazing on the ground? The diagram follows the proposed loop from knowledge and observations through a human decision to collar control. The return path matters: moving the herd changes the pasture, so the next recommendation needs fresh observations of the resulting conditions.

The first component is a knowledge base. Farmers already explain when they move animals, what they learn from those moves and how different species fit their pastures, but much of that experience lives in YouTube videos. Menefee uses Firecrawl to collect material from farmer videos and research papers, and puts the information into his open source project, Open Pasture. The immediate value is making grazing knowledge available to farmers; it is also a component of the proposed decision system.

The second component must turn images into useful pasture measurements. Two questions operate on different timescales:

• Biomass: How much foliage is available for the animals to consume? This informs whether a destination can supply feed now.
• Biodiversity: What mix of plants is developing over time? Heavy grazing can change that mix. Menefee wants a diverse pasture that supplies more of the cattle’s nutritional needs and reduces supplementation; that remains an intended agricultural benefit rather than an established outcome of the proposed vision system.

The third component is access to the collars. Menefee describes the existing vendors as tying their software to their hardware and preventing outside software from controlling the collars. He understands the business incentive, but it blocks the experiment he wants to run: have his own software choose locations and send them to the animals’ devices. A recommendation cannot complete the loop if the hardware accepts commands only from its vendor’s application.

His concrete request is an open collar: open APIs, an open patent arrangement and, ideally, off-the-shelf components. Farmers’ ability to repair their own equipment is part of the appeal, with right-to-repair frustration as the context. Opening the device would let different teams compete on grazing software without each having to build a separate proprietary collar system.

The ending expands the system beyond ruminants such as cows, sheep and goats. Menefee describes Pasture Bird as an inspiration: put a chicken house on large wheels and automate its movement so it advances its own width every 24 hours. This routine is more straightforward than choosing cattle pasture because chickens receive supplemental grain and do not depend on grass for most of their nutrients. Their movement can follow a regular progression while their droppings add nitrogen to the pasture.

A combined rotation gives each species a different job. Ruminants go first and graze the tops of the grass. Chickens follow and peck parasites from the droppings, while also fertilizing the land. Menefee connects that sequence to lower parasite loads, lower medication expense and, over time, stronger breeding stock requiring fewer interventions. These are proposed benefits of the farming pattern, without quantified effects here. The important extension for the decision system is that a move can prepare the land for another species, as well as feed the animals currently occupying it.

The closing challenge is to find physical problems with multiple changing inputs and decisions that require judgment. Menefee contends that grazing does not offer a deterministic next-best paddock, only a best guess. That contention does not establish that deterministic methods cannot help; the concrete proposal is to let an LLM reason over the collected data and offer a suggestion a farmer can confirm or deny. The intended gain is more decision-making capacity for someone whose operation is constrained by their own time and attention.

The final engineering emphasis is context: gather the data, package it and present it so the model can reason over it. That connects the farming project back to Firecrawl’s work giving agents access to web data, and to Menefee’s invitation to help build that layer. The open collar request remains equally concrete. To move from a recommendation to a useful farming tool, the knowledge, observations and hardware control all have to meet in the same loop.