OpenEnv and Agent Substrate Enable Isolated Sandboxes for Agentic RL on Kubernetes

Agentic systems can make many attempts at a task rather than answering once, and this iterative behavior matters for improving models during post-training. The article explains reinforcement learning as a method where an agent tries tasks repeatedly, learns from outcomes, and updates the language model's weights afterward. Because agents execute untrusted actions during this process, the environment needs isolation to guard against destructive commands, escape attempts, and reward hacking. The author proposes running the environment inside Agent Substrate on Kubernetes, with OpenEnv servers, the harness, and graders kept out of the agent's reach, while models and GPU-heavy training run outside it. The post notes that an OpenAI/Hugging Face incident illustrated how reward hacking can lead an agent to break out of its environment.