LLMs have a limit to how deep they can understand and refactor architectural issues.
That limit is far, far lower than a human's.