Post

I'm still waiting for an OSS project fully maintained by agentic loops that isn't a complete slopfest. Everyone can generate code, not everyone wants to maintain said generated code.

i don't think this will ever work. open source isn't just about generating code, it's about understanding the project's philosophy. maintainers don't just accept features because they sound good, they make decisions based on what the project is trying to be. i don't see an ai doing that.

i’ve tried this a couple of times. even when i explicitly tell it to keep the api as minimal as possible, it still ends up adding flashy features that are either out of scope or just over-engineered

Telling it to 'stay minimal' competes with a training bias toward doing more, and more usually wins. What's worked for me: show it actual rejected PRs from the project's history instead of a philosophy statement. Concrete no beats abstract yes.