Postmortem drafts

Postmortems get written late or not at all, because assembling the timeline takes two hours and everyone is tired. The assembly is mechanical. The analysis is not, and the agent must not pretend otherwise.

The job on one screen

Runs when	When an incident is marked resolved.
Reads	Incident channel messages, alert times, deploy history, ticket links.
Decides	What belongs in the timeline, and which facts are still missing.
Produces	A draft with a factual timeline, an impact statement, and empty sections for analysis.
Stops when	Draft posted with the gaps marked. Root cause is a human section, always.
Tools it needs	Chat history, monitoring, deploy history, doc API.
Autonomy to start at	Approve. It does the work and stops before the irreversible step.

Optional: load a real model

Run it

The checks the draft has to pass

The model writes, then code checks. A draft that fails any check never reaches the send step, and the failure is shown rather than silently patched.

Check	What it protects against
No individual named	A postmortem that names a person teaches everyone to hide problems, and you lose the reporting you depend on.
Impact stated in customer terms	500s on the export endpoint is not impact. Eight percent of exports failed for forty minutes is impact.
Duration present	Without a duration nobody can weigh this incident against the next one.
No root cause asserted	The draft states the sequence. Deciding the cause is the meeting, and an agent that pre-empts it narrows the investigation.

Where this one goes wrong

Naming people

The fastest way to make your incident process worthless. Check for it mechanically.

Asserting a cause

The first plausible cause is right about half the time, and the draft makes it the anchor for everyone who reads it.

Timeline from chat only

People discuss in chat minutes after events. Anchor the timeline on machine timestamps and mark human recollection as such.

How you would know it is working

Measure	Why that one
Postmortems completed within a week	The number this exists to move.
Actions created and closed	A postmortem with no closed actions was theatre.
Timeline corrections	How often humans fix the assembled timeline. Falling means it is getting trustworthy.

Earning more rope

Assist	Assemble the timeline only, a human writes everything else.
Approve	Draft the summary and impact sections too, with analysis left empty. Right level.
Auto	Never auto-publish. A postmortem is an organizational artifact, and it needs an owner's name on it.

Assemble facts, leave analysis empty. An agent that fills in the why turns the postmortem meeting into a review of the agent's opinion.

Related: Incident response · Alert deduplication · Multiple agents · all agent jobs · Agent Lab home

Free from AI School - no signup, everything runs in your browser.