Modal engineer Jason Mancuso demonstrated a practical recipe for training software agents with reinforcement learning. The system stabilizes mathematical training drift, compresses weight updates by a factor of one thousand, and cuts rollout compute costs by fifteen percent.