Heuristic versus LLM Agents in Agricultural Simulation
A distributed simulation testbed for precision agriculture, built to answer a focused question: do large language model agents make better farming decisions than classical rule-based ones, and at what cost? The system models soil–plant dynamics under stochastic weather and pits the two agent families against each other on the same fields.
Architecture
The system is deliberately decoupled into three services:
Simulator — the authoritative physics engine. It maintains global state (water balance, nutrient cycles, plant growth) and exposes it through an API.
MAS — the agents that perceive the field and act (irrigation, fertilisation, harvest). Two interchangeable kinds: heuristic agents with rule-based BDI-lite logic, and LLM agents that reason with OpenAI or Anthropic models.
Visualizer — a Next.js / TypeScript dashboard for live field state, agent logs, and yield metrics.
The physics engine
Soil–water dynamics are driven by realistic models rather than toy rules. Reference evapotranspiration uses the Hargreaves equation,
ET_0 = 0.0023 \,(T_{\text{mean}} + 17.8)\, \sqrt{T_{\max} - T_{\min}} \; R_a ,
with R_a the extraterrestrial radiation, while drainage follows a power-law in soil moisture. Plant stress is tracked as separate nitrogen- and water-stress factors that feed back into growth.
Evaluation
A dedicated harness runs multi-episode simulations and produces statistical reports comparing the agent types on:
Yield and water-use efficiency (yield per unit water),
economic return, and
AI token cost — so the LLM agents’ decisions are judged against what they cost to produce.
This makes the heuristic-vs-LLM comparison an honest one: the language models are held to the same agronomic outcomes and charged for their reasoning.
Source
Implementation and preset scenarios are in the source repository.