GitHub just shipped native Stacked PRs. CodeRabbit analyzed 470 pull requests. AI-authored ones carry 1.7x more bugs. 75% in logic and correctness. The monolith diff was always the wrong unit of review.
Claude Code writes a feature branch. 800 lines. One PR. The reviewer skims past line 300 and catches nothing at line 340 where the logic bug sits.
GitHub Stacked PRs (only available in private preview) split that into layers. Each PR targets the branch below it. 50-100 lines per diff. Cascading rebase. Each layer reviewable before the coffee gets cold.
It makes me realize that the review bottleneck might not be about speed, but diff size. I shipped an AI code review tool last year because Claude Code's output outpaced my review capacity. The tool helped because it would review line-by-line. Smaller diffs helped more.
Google and Meta ran stacked workflows internally for years with "Phabricator" and it exists similar tools if you just type "stacked pr" on google.
If AI writes 10x the code, more reviewers won't save you. Smaller diffs might.