Tutorial: loop an agent until the work is great

One prompt rarely gives you the best an agent can do. A Loop makes the agent its own critic: work, score, improve, repeat — until the result clears the bar you set.

1. Start a loop in one sentence

In the command bar, type a goal with a bar:

“loop: polish the landing page until score 9” — that’s a goal (“polish the landing page”), a bar (9/10), and Mastery HQ picks the top agent window to run it in.

2. Watch the passes

Open the Loop pane (Tools menu) to watch it think:

Each pass shows the work, the critique, and the score. You’ll see scores climb — 6, then 8, then the bar clears and the loop ends with its best version.

3. Tune the pressure

• Raise the bar: “…until score 9.5” for high-stakes work.
• Change the critic: add roast (“…until score 9 roast”) for a merciless reviewer, or karpathy for engineering taste.
• Cap the passes: loops stop at their pass limit even if the bar isn’t met, so they never run away.

4. When to reach for a loop

Anything with a quality gradient: landing-page copy, refactors, test coverage, a README, pitch decks. If you’d naturally say “make it better” three times, a loop says it for you.

Where to go next

• Loops: refine until good — scoring, lenses, and limits in depth.
• The Loop pane — reading passes, stopping, history.