ProRL-Agent-Server

Why we included this project

Teams doing reinforcement learning on real agent harnesses usually discover the harness was never built to be an RL environment. Polar lets you keep that harness as-is and treat it as the environment, then moves the expensive rollout work behind an HTTP server. The server dispatches requests to distributed gateway nodes that spin up runtimes, run the agent, build trajectories, and evaluate them asynchronously, so the training loop just sends requests and reads back results. It's trainer-agnostic and pairs with either vLLM or SGLang for inference, which means it drops into an existing stack instead of dictating your whole training setup. If you're scaling RL for coding or tool-using agents and want to cut GPU hours spent on idle runtimes, this is worth a look.