Ask HN: How do you get AI loops to converge?

For any arbitrary coding task, how can a Ralph loop break out of a local minima and converge, such that, even if the loop runs again, there are no tasks left to complete? These tasks are somewhat non-deterministic due to cross-interactions which make their stop conditions hard to define ahead of time (e.g a complex GUI)

I want to learn more resources for loop engineering, deterministic tooling etc.

What I am noticing is the agents/subagents oscillate between repeatedly patching the same bits of code over and over, or introducing new bugs.

Thank you!