Intel / video

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

Why it matters

Gives a specific stable recipe for asynchronous RL on your own traces and evals, and a cost lever from spreading rollouts across GPU types.

Key takeaways · AI-distilled

Jason Mancuso argues that teams with agents in production already have what reinforcement learning needs, a , evals and real traces, and that RL generalizes better than on recorded traces.
His stable recipe for asynchronous RL builds on GRPO with masked importance sampling, score centering and top-p sampling.
Sparse weight updates let Modal spread rollouts across clouds, GPU types and levels while training stays stable, and Mancuso says that makes rollouts 15% cheaper.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.