In my experience large scale automated refactoring of code is something that works reliably for the last 3-4 months.

But if you are saying that a human can instruct ai agents to refactor to prevent the big ball of mud, then you are saying that clean code *is* important.