my coding agent kept making the same stupid mistake over and over
like it knew how to fix it
but just... didn’t remember
it would:
- fail
- try something
- fix it
- then hit a similar issue later and repeat everything again
so I tried something simple:
→ when a fix works, store it as a pattern
→ next time a similar failure shows up, just reuse it
this already cuts a lot of loops
but now there’s a weird problem:
sometimes it overgeneralizes and applies the wrong fix in the wrong place
feels very human tbh
now I’m stuck between:
- not forgetting
- vs not overfitting to past failures
anyone else run into this with agent loops?
[link] [comments]