Recipes¶
Complete studies you can run, copy and adapt. Each is a file in the repository's
examples/ folder and runs with --dry-run (a fake model, no cost) or with a
real model via --model.
| Recipe | Shows | File |
|---|---|---|
| Negotiation | two agents, private information and reasoning, surveys before/after, a judge | examples/configs/negotiation.yaml, examples/negotiation.py |
| Support agent vs. a simulated user | user simulator, tools acting on a per-run world, policy checks | examples/configs/support_agent.yaml |
| Memory experiment | one agent across sessions, five memory types as conditions | examples/memory_tasks.py |
| Survey panel | persona-conditioned survey answers from token probabilities | examples/survey_panel.py |
| Social feed | a sampled population on a mini social network, clock, ranking as a condition | examples/configs/social_feed.yaml, examples/social_feed.py |
| A small town | Generative Agents-style residents: reflection, plans, impressions, a planted memory | examples/town.py |
| Your own environment | a repeated public-goods game written from scratch | examples/public_goods.py |
Real-model outputs quoted on these pages come from small open models (Qwen3.5-4B) with few seeds; they illustrate what the recipe measures, not findings.