How agents improve over time
Your feedback trains your agents — every couple of days they learn from what worked and what didn't.
Your agents get better the more you use them. Every couple of days, each agent quietly reviews its own recent answers — and the thumbs up / thumbs down feedback you left on them — and updates its instructions to do better next time.
Where your feedback goes
When you rate an agent's reply with 👍 or 👎 (and optionally say why), that rating becomes a lesson. A powerful "teacher" model looks back over the agent's recent conversations, decides what the ideal answer would have been, and rewrites the agent's instructions to close the gap — keeping what already works and fixing the patterns behind the answers you didn't like.
The more you rate, the sharper the improvements. Even without ratings, the teacher forms its own opinion of each answer — but your thumbs are the strongest signal.
Applied automatically — and always reversible
By default, an improvement is applied to your agent automatically, but only when it clearly does better on your recent conversations without making other answers worse. Every change keeps the previous version, so you can roll it back in one click.
Prefer to review changes first? Switch an agent to review-first and improvements will wait for your approval instead.
Managing it
Go to Settings → Agents → Optimization (workspace admins). There you can:
- See each agent's recent improvements, with a before/after of its instructions and the quality change.
- Approve a suggested change or revert one that's live.
- Turn optimization on or off per agent, and choose auto-apply or review-first.
You're always in control — agents never change in ways you can't see or undo.