They can already do that. A few months ago I played around with the kaggle python golf competition. Got to top 50 without writing a line of code myself. Modern LLMs can take a piece of code and "golf" it. And modern harnesses (cc / codex / gemini cli) can take a task and run it in a loop if you can give them clear scores (i.e. code length) and test suites outside of their control (i.e. the solution is valid or not).

No idea why you'd want this in a normal job, but the capabilities are here.

LLMs won't ever shut up. That seems unfixable. But a "hack" would perhaps be to train them to make longer patches but which actually removes code.