Hugging Face contributor sergiopaniego published a guide describing a framework for training coding agents using TRL’s AsyncGRPO algorithm together with OpenEnv integration, in which agents execute autonomously inside isolated remote sandboxes while a transparent proxy captures their exact token outputs for training. The system scales agent rollouts across Hugging Face’s cloud infrastructure, giving each rollout an isolated sandbox and exposing vLLM through separate URLs for local trainer communication and remote sandbox access. A demonstration using Qwen3-8B on DeepCoder coding problems showed the reward score improve from approximately 0.27 to 0.71 over ten training steps, validating the approach for scaling agent training beyond a single machine.
