[ForceBru]

I guess the human wants to "make existing, excellent code better". How to do this en masse? Make an LLM do this for them. It's well known that _sometimes_ (somewhat often, actually?) LLMs can indeed improve code (which makes sense: code is language, they're Large _Language_ Models, so "understanding" and (re-)writing text is what they do best), so it why not try to improve everything everywhere all at once?

One obvious reason is that if the LLM produces tons of garbage, this will waste the efforts of human reviewers. But if it's not tons of code _and_ the LLM wrote meaningful tests that pass (the existing tests must pass too), then the existence of such an agent (that only works with code and doesn't go off the rails writing blog posts etc) seems somewhat appealing.