Was there experience of fully delegating AI development to an agent without code review, and how did you systematically prevent issues like breaking neighboring components?