Good read, and I can relate to it. Agents tend to be great at the mechanical pass or the review, style or obvious bugs or missing tests, the part that already had linters, the part that doesn't delegate is intent, "Does this change mean what the PR says?"

Agentech review raises the floor on how many PRs we can file.But the human review is still the bottleneck here.

I believe that agents can never be the true only one reviewer to stamp a PR with an LGTM stamp. But they can definitely help a human.The human should still be the one who takes responsibility and ownership of the code he ships, but the agent can have the reviewer separating the important stuff from the noise in order to make reviewing code faster and easier. (disclosure: I work on a PR reviewer, Pyor Review, I know the pain well as a web dev so I stare at this a lot.)