Google GKE Labs Donates OpenRL to CNCF to Simplify AI Training
Google's GKE Labs has contributed OpenRL to the CNCF, aiming to untangle reinforcement learning infrastructure from AI research workflows on Kubernetes.
Reinforcement learning has emerged as one of the most resource-intensive operations in modern artificial intelligence development. Combining model training, inference, and sandboxed agent tasks into a single continuous loop, these workloads frequently intertwine complex research code with underlying computing systems. To address these infrastructural hurdles, Google LLC's Kubernetes engineering group developed OpenRL and recently donated the project to the Cloud Native Computing Foundation.
According to Sunil Arora, senior software engineer at Google Cloud, reinforcement learning functions as a comprehensive "super app" of artificial intelligence. Speaking in an interview at KubeCon + CloudNativeCon North America, Arora explained that the workload demands everything from active model training and data sampling to isolated agent execution and evaluation. This diversity continually tests the limits of existing cloud infrastructure.
Separating infrastructure plumbing from research loops
Initially launched as a research preview, OpenRL provides a self-hosted training API designed for deployment directly within an organization's existing Kubernetes cluster. The framework separates the reinforcement learning loop from the underlying compute plumbing. This architectural division allows infrastructure engineers to manage scaling and resource allocation independently, while AI researchers retain full control over their experimental parameters.
By abstracting the infrastructure into core primitives, teams can isolate the complexities of distributed computing from the actual machine learning logic. This separation is particularly vital when dealing with agent execution and reward calculation, which often requires running millions of isolated, secure commands in parallel.
To maximize efficiency, Google implemented an Agent Substrate component capable of packing a higher density of secure sandboxes onto constrained compute resources. Google developer Justin Santa Barbara noted that the need for robust isolation became apparent during the development of tools like kubectl-ai. Because autonomous agents possess the capability to execute potentially destructive commands, running them inside secure, disposable sandboxes ensures that any critical system errors remain entirely contained.
What it means for developers
For developers and AI engineers building complex agentic workflows, managing the infrastructure required for reinforcement learning can be a massive operational bottleneck. Projects like OpenRL aim to streamline these deployments on Kubernetes by establishing cleaner boundaries between experiment code and compute orchestration. For developers looking to experiment with leading models during their development cycles, platforms like Apixo allow you to try top AI models cheaply through one API at https://apixoai.online. As cloud-native ecosystems evolve to handle multi-step agent tasks and heavy fine-tuning jobs, standardizing these sandboxed environments will be essential for stable production deployments.
Source: Google’s GKE Labs turns reinforcement learning into a Kubernetes workload — SiliconANGLE AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
