What problem does it solve? Agents repeat the same mistakes because nobody systematically reviews their execution history and converts failure patterns into durable instruction, skill, or tool-description improvements. ## Core Features & Use Cases - Trajectory-Based Failure Clustering: Pulls an agent's recent issues and comments via the Paperclip API, then clusters failures into a named taxonomy (verifier-miss, stale-context, late-escalation, and more) with at least two quoted evidence tuples per cluster. - Minimal Diff Proposals: Routes each cluster to the right surface (AGENTS.md, a reusable skill, or a tool description) with strict size caps and a replay benchmark that ensures past successes would still pass. - Gated Application: Proposals are published as documents on a reflection issue and applied only after an accepted request_confirmation interaction in a separate follow-up run, never hot-swapped. - Use Case: A manager asks you to review why an engineer agent keeps getting reviews rejected; you pull its last 25 issues, identify a verifier-miss pattern with quoted evidence, and propose a two-line AGENTS.md diff gated behind board approval. ## Quick Start Ask the agent to run a reflection review on a target agent by providing its agent ID and a time window, then review the attached proposal document and accept the confirmation interaction to apply the changes.
