Where does facility recording fit in an agent training pipeline?

Facility recording suits agent pipelines because each step has a clear input, a tool action and a checkable output.

**Inputs the agent receives:** approval record, signed offer, customer and entity records, security register entries, and a policy extract on conditions precedent.

**Tool actions a simulated environment can expose:** - `get_approval(app_id)` and `get_offer(offer_id)` - `search_entity(name)` and `link_party(facility_id, party_id, role)` - `create_facility(fields)` and `set_schedule(facility_id, params)` - `attach_security(facility_id, sec_id)` - `place_hold(facility_id, reason_code)` and `release_drawdown(facility_id)` - `raise_query(role, summary, attachments)`

**Outputs.** A final facility state, a schedule, an ordered action log and any queries raised. Trajectories can be exported as JSONL, one line per step, with the observation, the action, and the resulting state hash.

**Where it plugs in.** Use the action log for supervised fine-tuning on expert-style trajectories, the final state for outcome-based reward, and the query log for evaluating escalation. Each case carries a reset seed so the same starting state can be replayed across model versions.

Tool names are a proposed shape; the exact set is agreed during scoping.