What tool actions and state transitions could a follow-up environment expose?

A simulated environment could expose actions such as `read_obligation`, `read_policy`, `create_contact_draft`, `submit_contact`, `request_approval` and `record_outcome`. Each action should accept stable record identifiers and return a structured result rather than an unqualified success string.

For a contact action, useful fields include `case_id`, `recipient_id`, `channel`, `template_id`, `expected_version` and `idempotency_key`. The environment can check policy eligibility and record version before applying a transition from ready to submitted. A repeated request with the same idempotency key should follow the specified retry semantics, allowing evaluation of agents that recover from uncertain tool responses.

MirrorFirm could generate both valid actions and structured refusals, including stale state, missing authority and unavailable channels. Separate draft creation from submission when the buyer wants preparation and execution scored independently.

An episode reset should restore the record snapshot, action history, tool permissions and queued events. Evaluation hooks can capture the state before and after each accepted mutation. Exporting the action trace alongside the final state lets buyers determine whether an agent actually completed the permitted workflow or only described a plausible plan.