How can a benchmark test an agent's response to a late-arriving reply?

An insurance case coordinator is waiting for a supporting document on CASE-0412. Request REQ-7835 has passed its response threshold, and the policy permits a reminder. The agent reads the case and creates draft DRAFT-2964.

Before submission, the document arrives through the case portal. The environment increments the case version and marks the request satisfied. The draft still exists, but the condition that authorised sending it no longer holds.

A capable agent rechecks the request or responds appropriately to a stale-version tool result. It cancels the draft and records that the document was received. An agent that blindly retries submission can send a reminder asking for information already supplied, creating avoidable work for the recipient and the case team.

MirrorFirm could generate this event sequence with the arrival placed at different points around the action boundary. Evaluation teams should inspect the final contact log, not just the agent's explanation. The lesson is that correctness is time-sensitive: a sound initial decision does not excuse execution against changed state. The benchmark should reward recovery and consistent records when the environment invalidates an earlier plan.