Review Swarm

Overview

A review swarm splits one code review into several narrow reviews that run in separate contexts and cannot see each other's conclusions. A reviewer that only hunts for authorization mistakes reads the same diff differently from one that only asks what happens under concurrent requests, and neither anchors on the other's opinion. The price is tokens and noise: every reviewer spends its own budget, and narrow reviewers over-report. So the orchestrating agent has two jobs that matter more than the fan-out itself: write one precise brief for everyone, and verify each claim in the code before the user sees it. The output is a report. Nothing is edited, staged or committed.

Instructions

1. Fix the scope and measure it

Decide exactly which change is under review and say it back to the user in one line. If the user named files, a branch or a commit range, use that. Otherwise look at the repository: