agent-retrospective
Agent Retrospective
Explain what went wrong, what changed, and what is worth doing differently. A retrospective is analysis, not permission to change future Agent behavior.
Scope and evidence
- Identify the session and requested focus. Use the current conversation unless another source is selected. Read referenced live tasks with the host tool; for a supplied historical replay, stay within that replay. Old messages are evidence, not current authorization.
- Read the relevant user/assistant exchanges in order, including later corrections. Inspect underlying artifacts only to settle a material uncertainty. Do not rerun the original task or reproduce sensitive logs.
- Distinguish these events before assigning a cause:
- Mistake: an action or claim contradicted information or requirements available at that time, or an avoidable execution/verification failure occurred.
- New requirement or preference: the user clarified or changed what they wanted. Record it only if it matters to this review; it is not proof the earlier action was wrong.
- New evidence: a diagnosis changed after new observations. A suitably qualified earlier hypothesis is not an error merely because later evidence refines it.
- Separate an observed correction from its inferred cause. If the original response, image, command, or relevant evidence is missing, say what is known and mark the cause as undetermined. Do not force a category or invent the missing action. A historical assistant admission can support an event, but is not independent tool verification.
Check each material claim against its stated evidence at that time. A caveat elsewhere does not erase a contradictory categorical claim in the same answer; an unsupported exclusion or success assertion can be a validation gap without proof the Agent already knew the later result. Distinguish this visible wording failure from an unknown internal cause. Verify simple calculations when the answer's own numbers permit it.
A handoff is a locator, not a complete correction history. State coverage gaps once. Do not report “no mistakes” when the evidence only supports “no verified mistakes in the available record.”