Pull request queues grow faster than reviewer bandwidth. On an active team, a developer might open three PRs in the time it takes a reviewer to work through one. The result is a stale queue, blocked merges, and context loss on both sides when review finally happens days later.
The deeper problem is that PR review bundles two distinct jobs: mechanical verification (does this compile, do tests pass, are there obvious bugs) and judgment (is this the right approach, does this fit the architecture). Humans are needed for judgment. The mechanical layer can be handled faster.
Fleet separates the mechanical layer from the judgment layer. A qa-engineer agent runs automated checks on PR creation. A tech-lead agent handles architectural judgment. Both subscribe to the same pr_needs_review event and run in parallel.
# Start the watcher — it handles label-to-agent routing automatically
fleet watcher start
# Check which agents are active
fleet status
The watcher daemon polls GitHub labels every two minutes. When needs-review appears, it matches the event to subscribed agents and starts them in tmux sessions. Each agent runs Claude Code with a role-specific prompt.
