Loop Builder

Most AI does the task. A loop does not stop until the result passes.

You do not hand it a task, you hand it an outcome. It makes an attempt, a separate critic scores that attempt against your goal, the gaps go back in, and it tries again. It stops when the result genuinely passes your bar, or it stops and tells you it is stuck. It is never allowed to simply declare itself done.

Watch one run

Starting a run...

Quality bar: 85 · Cap: 4 rounds

The five parts of a loop

The outcome

The finished state that ends the loop. Checkable: a number, a threshold, a pass or fail signal. "Write me a deploy script" is a task. "The next deploy ships in a day with zero rework" is an outcome.

The success criteria

The rubric the judge scores against. What has to be true before "passed" would mean you would actually ship it. Vague criteria make a weak loop, every time.

The actor

What produces each attempt, and the context it loads. It never grades its own work. That separation is the whole trick.

The judge

A separate critic, usually a different model. It returns a score and the specific gaps. The gaps are the fuel for the next round.

The controller

The rules. Pass, ship it. Fail, feed the gaps back and retry. No progress two rounds running, stop and ask a human. It never spins silently.

●

Why it matters

An ordinary agent tells you it is done. A loop has to prove it against a standard you wrote, with a critic that did not do the work. That is the difference between output and a result.

Try it · build a loop (nothing runs, this assembles the definition)

The result you want, not a task

An outcome describes what is TRUE when it is done.

What "done" looks like: success criteria, separated with a semicolon

The critic scores every attempt against these lines.

Max rounds and quality bar

It stops early if an attempt clears the bar, otherwise it refines up to the cap.

✓ YOUR LOOP, ASSEMBLED

This is exactly what the runner receives. Executing it costs real model time, so requests are approved by hand.

The library

1 trained loop saved from real runs. A loop that passes gets kept, and future runs learn from it. The contents are the operator's private playbook.

Want to see one run for real?

A run costs real model time, so it is approved by hand. Describe the outcome you would want looped and it goes straight to the operator.

The outcome you would want looped

The thinking behind it

Every failure of an AI system in production comes down to the same thing: it decided for itself that it was finished. The fix is not a better prompt. It is structural. Separate the thing that does the work from the thing that judges it, write down what passing means before you start, and refuse to accept "done" from the only party with an interest in being done.

That is all a loop is. It is also why a loop can be trusted with work that matters, and a single-shot agent cannot.

Run the same loop by hand, in any chat

You do not need this system to use the idea. Paste these, in order, into any model.

ACTOR
Produce an attempt at this outcome: <outcome>.
Load only what is relevant. Do not grade your own work.

JUDGE (start a new chat, so it has not seen the attempt being made)
Score this attempt 0 to 100 against these criteria: <criteria>.
List every specific gap. Be skeptical. Do not be generous.

CONTROLLER (you)
Score past the bar? Ship it.
Below the bar? Hand the gaps back to the actor and run it again.
No progress two rounds running? Stop, and change the approach.