Google's GKE Labs donates OpenRL, a Kubernetes reinforcement learning training API, to CNCF

Google's GKE Labs team, part of the company's Kubernetes engineering group, has donated OpenRL to the Cloud Native Computing Foundation. According to Sunil Arora, a senior software engineer at Google Cloud, the team first built an AI agent for Kubernetes tasks and then a benchmark to grade models on that work.

OpenRL debuted as a research preview in June 2026 as a self-hosted training API for a company's own Kubernetes cluster. It splits the reinforcement learning loop from compute management, so infrastructure engineers can handle scaling while researchers keep control of experiments.

Sandboxes matter because reward calculation can involve running millions of isolated commands. Arora said Google's Agent Substrate fits more of these sandboxes onto limited compute. Justin Santa Barbara, a Google developer, said his kubectl-ai tool could in theory format a hard disk, which is why he wants agent actions run in a sandbox. Google Cloud sponsored the segment, which was recorded at KubeCon + CloudNativeCon NA.

Teams running reinforcement learning on Kubernetes now have a self-hosted training option that keeps the training loop apart from cluster plumbing. Agents that execute generated commands should run in isolated sandboxes, which Santa Barbara's example illustrates. The interview does not cover OpenRL's license, availability or pricing, so those details need to be checked in the project itself.