Start with observable outcomes
Write down the result that should exist after a successful task, and what must never happen. Example: correct order refunded, wrong orders unchanged, policy satisfied.
Define action boundaries
Specify every tool the agent may use, validation rules, timeouts, and failure responses. Keep live customer systems outside the development environment.
Make episodes reproducible
Create deterministic seeds, reset operations, and pinned fixtures. Record agent actions and tool output so failures can be reproduced and debugged.
Treat rewards as tests
Prefer objective state checks wherever possible. When you need model-based judgment, test calibration against manually reviewed cases. Watch for reward hacking.
Validate with a baseline
Run a simple baseline agent, a random policy where meaningful, and deliberately incorrect behaviors. If obviously wrong actions receive high rewards, the environment is not ready.
Try the blueprint generator to structure a task, action space, and verifier before writing your first environment.
Generate a blueprint →