Your First Environment

Run one agent K times against a small task set and score every attempt. The grid makes reliability visible: full rows are already mastered, empty rows need a different intervention, and partial rows are the learning zone.

Files

• basic.py — the smallest complete environment: typed output, tasks, CodeScorer, and a K-attempt grid.
• with_summary.py — reads the grid through the stable summary() mapping.
• with_fingerprints.py — inspects the environment and policy fingerprints stamped on a run.

When to use

Start here when one successful agent run is not enough evidence. The examples pair an easy anchor with chained arithmetic calibrated to produce disagreement on gpt-5.5; an all-full grid is not a useful reliability example.

Continue to _02_task_sets/ when tasks need ids, metadata, or a checked-in JSONL file. The live turn-by-turn reward loop is not part of this release; these environments perform verification and dataset generation.

Run

python cookbook/environments/_01_first_environment/basic.py
python cookbook/environments/_01_first_environment/with_summary.py
python cookbook/environments/_01_first_environment/with_fingerprints.py

Requires OPENAI_API_KEY. Every model call uses OpenAIResponses with gpt-5.5.