What task and tool formats can a set-up task pack use?

For agent work, the useful unit is a task with a starting state, an allowed set of actions and an expected end state. A task pack for set-up could be a JSONL file where each line describes one case.

A task line could carry: `task_id`, `prompt`, `initial_state_ref`, `allowed_tools`, `policy_refs`, `expected_end_state` and `scoring_notes`. The `initial_state_ref` points to a snapshot of the linked tables, such as the approval, party candidates and open tasks, so the case can be loaded without running earlier steps.

For a simulated environment, the tool surface could include:

- `search_party(name, dob)` returning candidate parties with match scores - `create_account(party_id, approval_id, product_code)` - `attach_condition(account_id, condition_id)` - `open_task(account_id, task_type, owner_role)` - `set_service_state(service_id, state)` - `submit_for_review(account_id)`

Each call writes an event to the log. A `reset(snapshot_id)` action restores state between runs, and an `evaluate(run_id)` hook compares final tables with `expected_end_state` and returns per-check results.

For onward licensing, ask for the snapshot format, whether the environment runs as a container or a hosted service, and what restrictions apply to your customers running it. Treat tool names as indicative until scoped.