AI Agent Failure Postmortem: The 6-Line Checklist for When Your Agent Breaks Something

An autonomous agent doesn't fail like a server. There's no 500 error and no pager — just a quietly wrong outcome you discover later: the ticket auto-reply sent to the wrong thread, the cleanup job that "cleaned" live data, the agent that ran its loop forty more times after the input stopped making sense. The postmortem shape is different too, and most teams improvise it. Here's the six-line checklist that covers what actually matters.

Why agent incidents break normal postmortems

A human incident has a moment of detection — someone notices and shouts. An agent incident often has none: every step it took looked successful in the logs. The review question isn't "when did it break?" but "how long did it run before anyone could tell?" That changes what you must preserve first.

The six lines — in this order

Scope — what did it touch? Every record, email, ticket, payment or file the agent acted on, with counts. Guess here and you'll find the rest by rumor.
Kill — is it still running? Disable the trigger before you analyze anything. A postmortem written while the agent keeps taking actions is fiction.
Log — is the agent's log preserved? Copy it somewhere immutable before the next run overwrites it. This is the agent equivalent of keeping the black box.
Trigger — which input started it? The specific input, prompt or schedule change that set it off. "It just went wrong" is not a trigger; it's an absence of logging.
Guardrail — what should have caught it? Name the exact check that was missing: a dry-run flag, a spend cap, an allow-list, a human approval above some threshold.
Replay — re-run it in dry-run. Same inputs, actions logged but not executed. If you can't replay it, you can't prove the fix — you can only hope.

Where solo operators drift

Fixing the prompt and calling it a postmortem. The prompt is one line of six. The guardrail is the one that stops the next occurrence.
No human checkpoint on irreversible actions. Anything that sends, pays, deletes or publishes belongs behind an approval gate — that's the whole lesson.
Running the agent with the same credentials you use for everything else. Scope its access to what the task needs, so the blast radius of the next failure is a folder, not your account.
No "stop" ritual. Decide now how you'll disable it at 2am — one command, one switch, documented where you'll look.

The 20-minute version

Put the six lines — SCOPE, KILL, LOG, TRIGGER, GUARDRAIL, REPLAY — at the top of your agent incident template today. The first time an agent misbehaves, that page turns a dizzy afternoon into an orderly one. The second time, you'll already have the guardrail that makes it a non-event.

If you run agents unattended, the full playbook — the kill-switch design, scoped credential patterns, spend caps and the daily check routine — is here: Agent Ops 24/7 ($19). The free one-page version of the first-thirty-minutes card is still the best starting point: The First 30 Minutes (free).