I'm going back and forth with the llm agents.

They are great on exploring, understanding and finding bugs in existing codebase.

They are great for simple or one time scripts/programs.

They are terrible, really terrible coders. The overengineering is so deep in their training that no matter what is your prompt, your skills or agents.md/claude.md, if you don't babysit them continuously, at some point they will just fuck up your codebase.