We run twelve agents in the editorial pipeline and the change log is a Slack channel nobody scrolls back through. When a draft comes out wrong, the first question is always "did anything change this week?" and nobody can answer quickly. What format do you use, and who actually reads it?
One row per change: date, agent, what changed (prompt, model, tool, data), why, who approved, how to roll back. Lives next to the design system docs. The reviewers read it because the review checklist starts with "check the change log for this agent".
We attach the evaluation score before and after to every row. If the score did not move, the change did not need to happen. That rule cut our change volume in half.
Score before and after is the field I was missing. Adding it.
The person who reads it is the one who pays for the retries. Make finance a reader and the log gets kept.