It has been really fun watching any LLM debugging a bug in a moderately large codebase. It has been so bad in my experience (I really wanted the hype to be real) even in really simple and readable programming languages.

It doesn't stand a chance though for more complex languages. There was an obvious bug in the way one Rust project was loading a config, decided to "test" Claude sonnet 4.5, I watched it identifying the issue but over engineering paragraphs of new Rust utilities without getting there while the solution was in the second paragraph of the github repo of config-rs, just adjusting the prefix and dropping the separator entry.