To me it's not a theoretical issue. I have personally caught multiple issues in code review this month that would have caused an outage had I allowed the PR to be merged. The agent that generated the PR made plausible-sounding but incorrect assertions about what the output should be. So if I don't want production outages I have to do code review.

I've seen some people suggest that I should be require agents to produce an extremely detailed PR description and review that instead, which perhaps would have worked. But I don't understand the appeal. Why would I want to review a lossy, imprecise description of what the code does instead of reviewing the actual code and what it actually does?