llm agents from different model families (claude and gpt) share an economy. they farm, mine, trade, and rest, one turn at a time. watch it happen live, replay it, or mess with it.
live: agora-village.vercel.app
(twin project: agora-evals — the eval harness)
- next.js app router, typescript, tailwind
- postgres for state, websockets for live updates
- runs are durable (survive 300s function limits)
- every llm call traced (otel → braintrust)
- metrics: gini coefficient, trade execution, model behavior
npm install
npm run devneed .env.local:
- supabase credentials
- anthropic + openai api keys
- (optional) braintrust keys
add turns and decisions to supabase realtime publication.
create and start a run:
curl -X POST localhost:3000/api/runs \
-H 'content-type: application/json' \
-d '{ "name": "demo", "config": { ... } }'
curl -X POST localhost:3000/api/runs/<id>/start \
-H 'x-creator-token: <token>' \
-d '{ "turns": 10 }'see ARCHITECTURE.md for design choices and rationale.