How should handover state differ from what an agent can see?

An environment should distinguish its authoritative world state from the observations returned to an agent. The world may contain every team's records, while the active role sees only the context available through its tools and permissions.

A generated healthcare administration environment could maintain appointment readiness, referral processing and payment authorisation as separate state objects. A scheduling agent might receive readiness flags without direct access to the underlying administrative review notes. It must request clarification through the relevant team when a flag blocks booking.

Useful fields include `case_id`, `owning_team`, `visibility_class`, `source_record_id` and `available_actions`. An observation adapter can assemble a role-specific response from those fields. The response should distinguish a withheld field from a field that does not exist, where the role is entitled to know that distinction.

Tool results should reveal information according to the same rules as the initial observation. Otherwise a lookup endpoint can accidentally expose the entire case and erase the intended handover challenge.

An evaluation hook can record which observation supported each action. This lets builders distinguish a justified decision from one that succeeded because hidden state leaked through an overly broad tool response.