Quickstart

Seven single-file examples that cover the whole arc — run an agent K times, score every attempt, read the grid, export what passed. No folder hopping: each file stands alone and runs end to end.

Start here if you want the shortest path from nothing to a working environment. Once a file makes sense, the numbered folders that follow go deeper on the same ideas, one option per file.

Files

• _01_first_env.py — the smallest complete environment: tasks, a scorer, K attempts, and the pass-rate grid.
• _02_export_sft.py — keep the attempts that passed and write a conversational SFT dataset.
• _03_tool_reliability.py — verify an agent that calls tools, and check the calls actually executed.
• _04_judge_rubric.py — grade with a rubric when code cannot express the check.
• _05_compare_models.py — run the same environment under two policies and read the difference.
• _06_drilldown_demo.py — inspect a single attempt: verdict, transcript, tokens.
• _07_support_triage.py — a realistic classification environment end to end.

Run

One file:

python cookbook/environments/_00_quickstart/_01_first_env.py

The whole folder, in order, with the shared runner:

python cookbook/scripts/cookbook_runner.py cookbook/environments/_00_quickstart \
    --batch --timeout-seconds 1800 --json-report /tmp/quickstart.json

Each file makes real model calls, so OPENAI_API_KEY must be set and a full folder run costs real tokens. --timeout-seconds 0 disables the per-file limit if a task needs longer.

Where to go next

• _01_first_environment/ — the same first environment, broken into one option per file.
• _06_learning_zone/ — why tasks the agent sometimes passes are the ones worth training on.
• _10_export_sft/ — the export path in full, including provenance.