What tool actions and state can a generated referral environment expose?

A simulated environment could expose read_account, get_routing_policy, create_referral, attach_case_note and check_referral_status. Each action needs an input schema, a response schema and defined effects on environment state. A natural-language claim that a referral was made should not change the case unless the corresponding tool action succeeds.

A create_referral request could require account_id, interaction_id, reason_code and requested_contact_channel. Its response could return referral_id, submission_status and validation_errors. Specialist acceptance can remain a separate transition controlled by the environment, preserving the distinction between submission and receipt.

The initial state can include account records, queue availability, user permissions and unresolved interactions. Resetting a run should restore those records and remove mutations from the previous attempt. A fixture_id and generation_seed can identify the starting condition without becoming fields the agent uses to infer the answer.

For a reseller, the licensed package should describe which tool contracts and state fixtures are included, and which execution components a downstream customer must supply. An environment can then be evaluated as a sequence of permitted operations with observable consequences, rather than as a transcript whose apparent success depends on persuasive wording.