How does set-up map onto an environment's entities, states and tool actions?

Set-up maps cleanly onto a small state machine, which makes it straightforward to wire into an existing environment harness.

A typical mapping:

- **Entities:** approval, party, account or service, obligation (mandate, consent, contact), task, notification - **States:** approved, in set-up, awaiting information, active, referred, rejected - **Transitions:** each tool call moves one entity along an allowed edge, and illegal edges return an error the agent must handle

Tool actions a release could expose include `search_party`, `get_approval`, `create_party`, `open_account`, `attach_obligation`, `set_status`, `raise_referral` and `send_notice`. Each takes record identifiers such as `APP-2187` or `ACC-5530` and returns the changed record, so a harness can log the diff after every step.

Preconditions can be encoded as guards. For instance, `open_account` could fail unless the approval status is approved and its expiry date is in the future. That lets you test whether an agent reads the failure and recovers rather than retrying.

For the evaluation hook, expose a read-only `snapshot()` that returns all entities as JSON, and compare it to an expected end state after the episode. Scope the exact entity list, status vocabulary and guard rules in a pilot, since they differ by domain and by the systems you want to imitate.