Google donates OpenRL to CNCF for Kubernetes RL workloads

Google's GKE Labs builds experimental projects and puts them in users' hands early. OpenRL emerged from that process, starting as a research preview in June 2026 before being donated to the Cloud Native Computing Foundation. The project's origin traces to a concrete internal need: Google built kubectl-ai, an AI agent for Kubernetes tasks, and then a benchmark to grade models on that work. As Justin Santa Barbara put it, kubectl-ai could in theory have formatted his hard disk, which made a sandbox a practical requirement rather than an abstraction.

The architecture reflects that. OpenRL splits the reinforcement-learning loop — which mixes training, sampling, agent work in a sandbox and evaluation — from the compute plumbing, so infrastructure engineers manage scaling while researchers keep control of experiments. Sandboxes are where the loop meets agent infrastructure, since calculating rewards can mean running millions of isolated commands. Google's Agent Substrate, per Sunil Arora, packs more of those sandboxes onto limited compute, addressing the density problem that arises when reward calculation is command-heavy.