What they did
The author, Adra, found that reviewers on their team were picked more or less at random: someone posted a PR in the group chat and waited. They used Cursor Automations to run an agent on each new pull request. The agent reads the change, builds a pool of people with relevant history in the touched files or nearby components, drops the author and bots, and then ranks that pool by outstanding review requests, reviews assigned in the last seven days, time since their last assignment, and only last by experience. It requests exactly one reviewer, leaves any human choice alone, never re-requests someone a person removed, and reports who it chose and why. It is not allowed to review, approve or merge.
How it went
The pilot started on 2 November 2026. Mean open-to-merge time that week was 12.9 hours, against 18.2 hours across the three weeks before. Against the week just before, the mean went from 25.8 to 12.9 hours and the median from 14.4 to 6.2 hours. The author stresses these are one week of observational numbers and does not claim the routing caused the whole drop.
Worth knowing
Ranking by expertise alone would keep picking the same person, which is why workload comes first inside the qualified pool. The prompt also tells the agent to skip any metric it cannot get and say so, and to flag a PR for human routing rather than guess when no one qualifies.