Which upstream and downstream records does an agent link together during set-up?

Most set-up processes can be modelled as a small graph. The approval is the root, and every record the agent creates should point back to it.

**Upstream:**

- `approval_id` (for example `APP-2187`), `decision`, `decided_at`, `decided_by_role` - `applicant` block with name fields, contact details and identity check status - `requested_product`, `conditions` and `required_documents`

**Downstream records an agent may create or modify:**

- `customer` with `customer_id`, `status`, `source_approval_id` - `account` with `account_id`, `customer_id`, `product_code`, `opened_at`, `terms_ref` - `service` or `entitlement` with `service_id`, `account_id`, `activation_state` - `task` with `task_id`, `assigned_team`, `related_ids`, `due_rule` - `audit_event` with `actor`, `tool`, `target_id`, `before`, `after`

**Links that matter:** `account.customer_id`, `service.account_id`, `task.related_ids` and every `source_approval_id`. A checker can walk these edges and flag orphaned records, missing links and records tied to the wrong approval.

**Tool actions** typically include `search_customer`, `create_customer`, `reopen_customer`, `open_account`, `activate_service`, `create_task` and `add_note`. Each should return the new identifier so the agent can pass it to the next call.

Export formats such as JSONL (one record per line, typed by a `record_type` field) or Parquet tables per record type both preserve these links. Agree the identifier style and referential rules in the pilot so the same graph checks run on every release.