When an agentic loop won't converge, what I had to fix was not the output but the frame I set at the start

Summary

When I have an AI agent do the same job over and over and what comes back never gets any better, what I should have fixed was not each individual output but the frame I handed the agent at the start. By frame I mean the wording of the request I give the agent, the material I have it reference, and the condition for how far it should go before it stops. If the same mistake keeps coming back across rounds no matter how I fix the outputs one by one, the root is on that side.

Here are the actual numbers from this loop. In a loop to raise the design quality of a site, I used the count of a reviewing agent’s findings as the stopping condition, and the count moved as 4, 2, 3, 2, 2, 1, 0, 2, 2, 0. After it fell once to 0, it went back up to 2 in the next round. That is, after a round in which the reviewing agent found nothing to fix, it found two more. This return was the signal to suspect the frame rather than the output.

This article is about how to switch from the reflex of fixing the output to suspecting the frame. Up to here, which you can read for free, I have stated the conclusion. In the body, which you can read with a subscription, I trace three typical ways the frame breaks, the signal for when to switch judgment, and how to decide when to stop, in the order of the failures that actually happened. What I write here is my own thinking, based on the operating records of agentic loops I actually ran.

Who this is for and what you can take away

This article is for people running a mechanism that has an AI agent do the same job repeatedly. When you hit a loop that will not improve no matter how you fix it, you can take away when you should stop fixing outputs and suspect the frame you handed over at the start, where that frame tends to break, and how to stop based on whether the findings have run out rather than by deciding the number of rounds in advance.