Orchestrates cloud coding agents in sandboxes that self-verify. Catches real bugs with 79.5% precision at ~$1 per review.