What tool actions and linked records does an escalation agent work with?

An agent working inside an escalation process usually needs a small set of tools over a handful of linked records.

Typical tool actions in a simulated environment:

- `get_case(case_id)` returns the application or review item with its status and assignee - `get_authority_matrix(role, product)` returns limits and conditions - `get_exposure(party_id)` returns linked facilities and balances - `escalate_case(case_id, target_role, reason_code, note)` changes status and assignee - `request_information(case_id, question)` returns the case to the requester - `record_decision(case_id, outcome)` is refused when the actor lacks authority

Linked records share keys. A `case_id` links to `party_id`, `facility_id`, `policy_id` and `reviewer_id`, and an `escalation_id` links back to the case and the reviewer it came from.

Each tool call can return a structured error such as `AUTHORITY_EXCEEDED` or `TARGET_UNAVAILABLE`, so an agent's reaction to a refusal can be scored. Environment state is seeded from a scenario file and can be reset between runs, which lets the same case be replayed against different agent versions.

Scope the exact tool surface in a pilot so it matches your own system's API, since connected systems vary.