The pull-request queue became AI's new bottleneck

AI code generation was sold as a productivity tool for engineers. The category it actually produced is a productivity problem for reviewers.

The bottleneck in this cycle did not move from engineers to machines; it moved from the keyboard to the queue. CodeRabbit's Loker describes the new shape of that queue in one line: Triage ranks each pull request across business value, urgency, risk, effort, readiness, dependencies, linked issues, ownership, and reviewer fit, then assigns it to a priority band. GitHub shipped a competing triage dashboard to general availability on July 9. The mechanism is the same on both sides: triage the flood, map the blast radius, security-scan before any human reviewer reads the diff.

CodeRabbit's own admission sharpens the stakes. Senior engineers reportedly rubber-stamp pull requests over roughly 500 lines, and architectural feedback is rare under file-by-file or alphabetical review. The agentic layer is being sold into that exact failure mode. The machine did not remove the engineer at the gate. It gave the engineer a much bigger queue and a new set of dashboards to defend the gate with.