I read only a small part of what agents write now, mainly risky chunks after 3 review passes by different model families. Reading everything dosnt scale. I think a lot of it comes down to the confidence you have in the development procedures for the agents, the system architecture of what they are changing and the review procedure.