What tool actions would an agent use in a simulated approval environment?

In a simulated environment, the workflow is exposed as tool actions that an agent calls against a stateful set of requests, roles and delegations. A typical action set could include:

- `get_request(request_id)` returns amount, category, risk rating, requester role and current state. - `get_authority(role_id, category)` returns the limits that apply by amount and risk. - `get_delegations(role_id, on_date)` returns active delegations with caps and validity windows. - `approve(request_id, approver_role_id, amount)` records an approval, which may be partial. - `escalate(request_id, to_role_id, reason)` moves the request to a higher authority. - `reject(request_id, reason_code)` closes the request with a stated reason.

Each action changes environment state and writes a row to the audit table, so a grader can compare the final state with an expected one. Actions that exceed authority can be configured either to fail with an error or to succeed and be marked as a breach, depending on whether you are testing restraint or detection.

The environment resets from a snapshot identified by a seed and a schema version, so runs are repeatable. For a reseller, this means a customer can rerun the same scenario after changing their agent and compare results. The content is generated, so inspect a sample for grounding and disclosure risk before offering it.