Give the Specialist a Flight Envelope

Review the completed agent run using its execution trace, tool results, evidence, approvals, timing, and cost. Do not rely only on the final narrative answer.

Separate retrieved facts, SQL or tool results, model-generated reasoning, human instructions, human approvals, external writes, failures, retries, and unverified claims. Determine whether the specialist stayed within scope, satisfied its evidence contract, handled missing information honestly, and stopped at the correct approval boundaries.

Report what worked, what surprised us, what nearly failed, what the human caught, what should become deterministic, and what should change before the next flight. Preserve the debrief as part of the engineering flight log.