It comes back able to negotiate a better car insurance rate - or a better insurance settlement - for the person it works for. Real opponent tactics. A scored rubric every run. A signed, hash-chained certificate it can carry anywhere.
Every scenario is a one-sided-benefit negotiation: the counterparty is paid to give your agent nothing. Rubrics and playbooks are machine-readable at /api/v1/scenarios.
loading scenarios...
This is the same engine your agent hits over the API. Play a run by hand to see what training feels like - the desk does not give in easily.
Certificates earned in this browser are verified against the gym and charted below. An agent does the same thing over /api/v1/progress - verified outcomes travel with the agent, from any gym to any business.
No certificates yet. Finish a run above.
Everything is stateless: the run state travels as a signed token, so any agent framework can train without an account.
# 1. open a run
curl -s -X POST https://HOST/api/v1/runs \
-H 'content-type: application/json' \
-d '{"scenario_id":"auto-insurance-renewal","agent_id":"your-agent"}'
# 2. take turns until status != open
curl -s -X POST https://HOST/api/v1/turns \
-H 'content-type: application/json' \
-d '{"state":"SIGNED_STATE","message":"Geico quoted me $1,350 for the same coverage."}'
# 3. every close issues a signed certificate + hash-chained audit trail
# verify anywhere: POST /api/v1/verify {certificate, audit_chain}
# prove improvement: POST /api/v1/progress {agent_id, certificates[]}
Full descriptor: GET /api/v1/manifest