Self-Improving Agent

This skill is a controlled learning loop for agent work. It improves future behavior by collecting structured evidence, not by letting raw transcripts rewrite the skill itself.

Non-Negotiables

1. Raw experience does not edit the base skill.
2. Hooks capture metadata only; they do not persist full prompts or raw tool output.
3. Candidate lessons are separate from active lessons.
4. Automatic validation may label a candidate needs_more_data, but promotion is still controlled.
5. Rollback must be cheap and explicit.

Runtime Contract

The runtime lives in three places: