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

Summary (tech + product concepts)

Cody (Firecrawl) argues that people—including engineers—are “not thinking big enough.” He uses automated pasture rotation for grass-grazing systems as an example of a large, real-world, data-heavy problem that needs AI agents + context, not another small SaaS MVP.

Problem: scaling grass-based animal husbandry

• Most cows are fed in feedlots; only a small fraction are on pastures.
• Goal: increase how many animals can graze on grass by improving rotational grazing.
• Core challenge: rotational grazing requires constant daily movement of:
  • animals
  • fences / “pens”
  • water access
  • and ongoing monitoring of pasture conditions
• Farmers must avoid degrading pasture by:
  • letting cows graze favorites too much
  • trample areas
  • managing grass height and growth stage (there’s an optimal “golden zone” of grass age)

Existing tech and its bottleneck

• Companies like Halter and NoFence use GPS collars with virtual geofences to remotely constrain animal movement.
• Cody’s critique:
  • These systems don’t solve the decision-making.
  • You must determine where to move animals based on changing conditions (e.g., drought, precipitation, grass regrowth patterns).
• Current pasture management often relies on human intuition and repeated on-site monitoring, including knowing grass height and timing.

Proposed solution direction: LLM-driven “autonomous grazing”

Cody suggests placing an LLM as the center of a decision loop that outputs the “next best” paddock/placement for:

• individual animals (where a cow should go next)
• the herd (best collective move)
• broader ecosystem/farm goals

To enable this, the system needs multiple inputs, such as:

• GPS locations of animals (and where they were yesterday)
• inferred/predicted movement options (where they might go tomorrow)
• local drought/weather conditions
• grass height / pasture condition across the farm

He emphasizes this is not deterministic:

• there isn’t a single “correct” answer
• only the best hypothesis that a human can verify

Data/vision options for grass monitoring

Cody outlines three approaches to measuring pasture state:

1. Drone orthophotos

   • Potentially high-resolution grass imagery.
   • Bottlenecks: farmer skill to operate drones, regulations, and limited ability to deploy fully autonomous drones in many jurisdictions.

2. Satellites (e.g., Planet)

   • High-frequency global imaging (claimed ~1m resolution).
   • Bottleneck: satellites are often too high to extract fine-grained details needed for grazing decisions.

3. Reference-camera approach (experimental idea)

   • Place a camera + measuring reference near a tree/object.
   • Use the scene to estimate relative grass height and growth progress.

System building blocks he emphasizes (engineering “components”)

Cody says the full solution needs three major components:

1. A knowledge base

   • Much grazing know-how lives in YouTube and informal sources.
   • Firecrawl helps collect and structure this info.
   • He references an open-source project: Open Pastures, intended as a repository so farmers can access this knowledge.

2. Visualization / mapping of pasture variables

   • Key variables include:
     • where animals are
     • where they should be moved
     • biomass/foliage availability
     • longer-term biodiversity and grass composition balance (cool-season vs warm-season grasses, legumes)
   • The aim is to reduce reliance on supplemental inputs (hay, minerals, etc.) through better pasture nutrition.

3. Geofencing / collar integration

   • Geofencing tech is helpful, but he strongly disagrees with vendor lock-in:
     • collars must be purchased from the vendor
     • you can’t easily connect your own software/stack to the collars
   • His preference: open protocols/APIs so innovators can plug in and improve decision logic.

Call to action: open collar + open APIs

He explicitly calls for community help to create:

• a collar with open patents
• open APIs
• ideally a ready-to-integrate solution that enables competition at the software layer (decision/optimization layer), not only hardware.

Extension beyond ruminants: multi-species pasture automation

Cody also teases that the approach generalizes beyond cows:

• Example: Pasture Bird
  • A mobile, automated chicken coop moved daily.
  • Rationale: chickens fertilize soil via droppings; ruminants + chickens can reduce parasite load and medication needs.
• Multi-species rotations could improve farm robustness and reduce intervention over time.

Takeaways (how he frames the “big” technical thesis)

• Choose real-world physical problems requiring multidimensional analysis (not deterministic yes/no outputs).
• Build data-intensive systems where an LLM makes hypotheses humans can confirm or disprove.
• A major bottleneck is context engineering:
  • collecting the right internet data
  • packaging it
  • presenting it so an LLM can reason effectively.

Product/company positioning

• Cody ties this to Firecrawl: building the “context layer” for AI agents using web data APIs.
• He also mentions hiring and directs viewers to job listings on the company website.

---

Main speakers / sources

• Speaker: Cody Menefee (Firecrawl development team; presenter of the talk)
• Referenced companies / sources: Halter, NoFence, Planet, Pasture Bird, University of Missouri (research assistant mentioned), plus grazing knowledge from various YouTube farmers (general category)
• Referenced projects: Open Pastures (Cody’s open-source project)