Agent Reflect Loop

Every agent team has the same scar tissue: the agent makes a mistake, a human corrects it, the agent apologizes beautifully — and makes the exact same mistake next week. Corrections evaporate because nothing in the system is designed to keep them. This skill builds the reflect loop: a disciplined cycle that captures corrections and failures, distills each one into a durable rule or check, regression-tests the rule against the original failure, and stores it where it will actually fire next time (prompt, skill, eval, or guardrail — chosen deliberately, not hopefully).

To give an agent system institutional memory: a repeatable process that converts each correction, failure, or near-miss into a tested, stored improvement — and verifies the improvement actually prevents recurrence. The output is both the loop design (capture → distill → test → store → verify) and its first full cycle run against the team's real recent failures.

• When the same agent mistakes recur across sessions ("we corrected this three times already").

• When onboarding a new agent or a new team member to existing standards — the rule store is the onboarding.

• After any incident or near-miss: the reflect loop is the postmortem's action-item engine.

• When building long-lived agents (weeks/months of operation) where statelessness means relearning everything.

• For one-off mistakes with no pattern — fix the instance, do not build a loop around it. The loop earns its keep on recurrence.

• When the "corrections" are actually disagreements about taste with no stable right answer — the loop stores rules, not preferences; it will faithfully learn your inconsistency.

• When no human reviews the distilled rules — an unreviewed rule store accumulates cruft and contradictions. The loop requires a reviewer.

• As a substitute for fixing a fundamentally unreliable agent — if it fails 30% of the time, the loop will drown in corrections. Fix reliability first.