AgroManager

AgroManager is a hosted OpenReward environment for long-horizon UK arable farm management under live uncertainty. The project prioritizes a deployed environment first: stochastic weather, stochastic prices, weather-driven pest pressure, and noisy soil drift all happen while the agent is acting, and the baseline path uses the actual hosted environment devatreya/AgroManager.

The environment models a 400-acre Cambridgeshire farm split into 4 plots over 40 quarters. The main metric is mean_terminal_score, which balances solvency and final soil stewardship. The strongest scripted baseline is weather_aware_rotation, which adapts to live quarter-by-quarter price boards and recent weather history while protecting weak-soil plots and buying irrigation only when it is economically justified.

Quickstart

python3.12 -m venv .venv312
source .venv312/bin/activate
pip install -r requirements.txt

python scripts/fetch_weather.py
python scripts/fetch_prices.py
python scripts/build_tasks.py --scale medium

Run the local server for smoke checks:

uvicorn app:app --host 0.0.0.0 --port 8080

Run the hosted smoke test:

export OPENREWARD_API_KEY="..."
python scripts/test_deployed_env.py --openreward-env-id devatreya/AgroManager

Run hosted baselines:

python eval/run_baselines.py --split train --capture-conversation --openreward-env-id devatreya/AgroManager
python eval/run_baselines.py --split validation --capture-conversation --openreward-env-id devatreya/AgroManager

Prepare SFT data from hosted rollouts:

python scripts/prepare_sft_data.py --openreward-env-id devatreya/AgroManager

Run Modal jobs:

modal run pipeline/train_sft.py
modal run pipeline/train_rl.py --openreward-env-id devatreya/AgroManager
modal run pipeline/eval_compare.py --split validation --max-tasks 16 --openreward-env-id devatreya/AgroManager
modal run pipeline/eval_compare.py --split test --max-tasks 16 --openreward-env-id devatreya/AgroManager

Project Layout

• app.py: canonical OpenReward server entrypoint
• env.py: hosted environment class AgroManager
• sim.py: simulator state, stochastic dynamics, reward logic
• baselines/: scripted policies, with weather_aware_rotation as the main benchmark
• pipeline/: hosted session cache, transcript compaction, rollouts, Modal training and eval scaffolds
• scripts/: data ingestion, task building, hosted smoke test, SFT prep
• eval/: hosted baseline runner
• docs/: design, deployment, training, and failure-mode notes
• data/processed/: processed Cambridge weather, DEFRA price anchors, and scenario tasks

Data Sources

• Weather: Open-Meteo archive, aggregated to Cambridge quarterly climate context
• Prices: DEFRA agricultural price indices, chain-linked from the legacy ODS series to the current CSV series

field_beans uses a documented historical proxy chain-link before the modern field_beans series begins. There is no synthetic weather or price fallback path in the runtime.

OpenReward Deployment

orwd create AgroManager \
  --description "Long-horizon UK arable farm benchmark with Qwen + Modal + ART" \
  --namespace devatreya

orwd link devatreya/AgroManager devatreya/AgroManager
orwd upload devatreya/AgroManager ./data/processed/

orwd deployments devatreya/AgroManager
orwd logs devatreya/AgroManager

Notes

• The acceptance path is the hosted environment, not localhost.
• The simulator remains stochastic inside the episode; deterministic replay is not the primary benchmark path.
• training/sft_result.json, training/rl_result.json, and eval/*_comparison.json are written by the corresponding entrypoints.