This is a very interesting idea and I really should consider Aider in the "scriptable" sense more, I only use interactively.
I might add another step after each PR is created where another agent(s?) review and compare the results (maybe have the other 2 agents review the first agents code?).