Just this week, I used Opus to build out a feature. It did a great job. I could've shipped it without reading the output and there'd have been no repercussions. Then I glanced at the code. It had been needlessly thorough in covering every possible edge case and wrote tests for scenarios that never would've happened. That code was resilient and well-written. But it was also rigid and removed from the spirit & purpose of the specific business need. Enter the next agent: it would've assumed that robustness was there for good reason, and would've gone to further lengths to preserve it as it modified or extended the feature for the current prompt. After a few cycles of this, the team would've ended up with a functional, rock-solid service performing as designed. But it also would've been out of alignment with the people it's intended to serve, slower to deploy, and more difficult to change as new needs arise. All this is only multiplied by the fact that it's built & maintained by multiple teams. You don't have to read every single line. Santiago says that. But good grief, resist the silly, growing suggestion that reading any code is a thing of the past.