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!
wrote some learnings here: https://uberblick.ai/blog/2026/built-on-its-own-pages
for GUI work the stop condition that worked for me is interaction + visual tests.
make sure all component interactions are automatically tested. screenshot the screens the task touches, and let your agent see what changes it made
We're doing something similar in DeepXplore, where a composer spins up agents as needed, checks their results and decides what to do next. We also cap the iterations so things don't run forever. Wrote a bit about how it works here: https://deepxplore.io/blog-entries/blog-deepxplore-code-agen...