Learn
What is an agent retrospective?
An agent retrospective reviews a period of agent operations — what shipped, what failed, correction categories, spend, and skill changes — to improve roles the way engineering teams run sprint retros.
In depth
Continuous operation without reflection repeats mistakes. A agent retro (weekly or monthly) asks:
- What did roles ship vs plan? - Where did humans override or reject output? - Which correction categories repeated? - Any incidents, near-misses, or budget surprises? - What skill or eval changes do we commit to?
Output: 1–3 concrete skill PRs, not vague 'prompt better.'
Link retros to accountability logs and evaluation trends — data beats opinions.
Founders running one-person companies benefit most: 30 minutes Friday beats discovering the same support failure every Monday.
Examples
- Weekly — support reject reasons → 2 new eval cases
- Monthly — cost per accepted task down 12% after model routing tweak
- Retro action — add escalation skill for billing edge cases
- Share retro notes in git next to skill folder
Related terms
FAQ
Retro frequency?
Weekly for new roles; monthly when stable.
Who runs it?
Role owner with optional ops lead; founder for solo companies.
Run governed agent roles on a company OS — not only definitions in a glossary.