How long do you see the humans in the loop being necessary?
When your AI-managed codebase breaks, who are you going to ask to fix it? The AI?
But if it does it could still fix it.
And you won't have to tell it anything, alerts will be sent if a test fails and it will fix it directly.