Vercel’s Post

35% of merged PRs is the headline, but the number I'd actually want next to it is median review time on the agent-authored ones. In my experience the bottleneck doesn't disappear when an agent writes the PR, it moves — reviewers still have to build the same mental model of the change, just without having written it themselves, which can slow review down for anything non-trivial.

The reviewer bottleneck is the number we'd want to see next to that 35%. In our own rescue work, agent-authored PRs usually pass CI cleanly, the failures we catch later are domain logic issues, a fix that passes tests but doesn't match how the business actually handles an edge case, on older codebases with rules nobody wrote down. Is your review headcount shrinking too, or shifting toward catching that kind of thing instead?

Can’t wait to harvest ideas for my own software factory. Thanks for sharing your process