The post describes a method for RL-training coding agents where the agent’s own harness owns its execution loop rather than being driven by the trainer, using TRL’s AsyncGRPO trainer integrated with OpenEnv to scale rollouts across isolated remote Hugging Face sandboxes (one container per rollout). A transparent proxy captures the exact tokens and logprobs the agent produced inside the black-box harness, and a verifier scores solutions against hidden test cases to compute rewards. In a demonstration training Qwen3-8B on DeepCoder problems over 10 steps across 32 problems, reward improved from about 0.27 to 0.71.
