Training Your Own Agent Models With RL Is Now Practical - Jason Mancuso, Modal

Why watch

Jason Mancuso from Modal explains why reinforcement learning outperforms fine-tuning for training agent models, and presents a practical recipe for implementing asynchronous RL using GRPO with masked importance sampling, score centering, and top-p sampling. The talk is for engineering teams already running agents in production who want to move from treating models as swappable backends to training their own custom agents, with guidance on spreading rollouts across multiple clouds and GPU types while reducing costs by 15%.