Agent Retrospective

Improve agent behaviour from observed friction, not borrowed prompt folklore. The output is an evidence-backed change proposal; editing starts only after the user chooses what to adopt.

1. Scope the sample

Agree on:

• repositories or working directories;
• harnesses and models;
• history locations;
• time range or number of sessions;
• any sensitive projects or content to exclude.

Discover history read-only. Do not scan unrelated machines, accounts, or directories merely because they are accessible. Record the task mix for each model and harness so difficult work assigned to one model is not mistaken for a model effect.

Done when: every history source and exclusion is explicit, and the sample can be reproduced.

2. Build the evidence set