Google Cloud hands Agent Substrate, gVisor and OpenRL to CNCF

Google has donated three AI-focused open-source projects — Agent Substrate, gVisor and OpenRL — to the Cloud Native Computing Foundation, with all three accepted into the CNCF Sandbox on 11 November 2026. The move places software for running and training AI agents under vendor-neutral governance at a moment when companies are managing growing numbers of automated software tasks on shared infrastructure.

The donation covers the full arc of the AI agent stack: a runtime for stateful agent sessions, a sandboxing layer for untrusted code, and a framework for reinforcement-learning post-training on Kubernetes. Each project arrives with concrete production credentials rather than empty repositories.

What is Agent Substrate?

Agent Substrate, the newest of the three, is designed to run stateful agent sessions without keeping a dedicated container active while an agent waits for a model response, an external tool call or human input. The software separates an agent's identity and stored state from the compute used to execute tasks, allowing paused sessions to be restored in under 500 milliseconds.

Drew Bradstock, VP of Product Management, Orchestration & Kubernetes at Google, said Agent Substrate can run millions of sandboxes at 10 times the density of standard container runtimes, with more than 500 suspend-and-resume activations per second. The project has attracted more than 100 contributors since public development began in May 2026, according to Bradstock.

The project has also reached version 1.0 and is generally available on GKE, Google's managed Kubernetes service. The managed GKE offering targets production deployments, while the open-source code can run on any Kubernetes cluster.

Agent Substrate runs standard OCI containers (the Open Container Initiative format used by Docker and Kubernetes) and can use either Cloud Hypervisor microVMs or gVisor as its isolation layer.

Why do agent workloads need different infrastructure?

Bradstock argued that agent workloads differ fundamentally from conventional cloud applications. They execute code in short bursts, pause frequently, and often need strict controls over tools, secrets and external services. Standard container runtimes, built for long-running services, waste resources on idle agents waiting for external responses.

That idle waste matters at scale. If a company runs thousands of agents that each pause for seconds or minutes while waiting on model responses or human approvals, dedicated containers sit unused. Agent Substrate's suspend-and-resume model, with the sub-500-millisecond restoration figure, is Google's answer.

The isolation question is where gVisor enters the picture.

What does gVisor bring to the CNCF?

gVisor, first open-sourced by Google in 2018, is a sandboxing technology that sits between a container and the host operating system. It intercepts and handles system calls in userspace, so code inside the container does not interact directly with the host Linux kernel.

That architecture matters for AI agent workloads because agents may generate shell commands or run unreviewed third-party packages, Bradstock said. Standard containers share the host kernel, raising the risk when code is untrusted.

Google said gVisor already runs millions of sandboxes a day inside GKE Sandbox and Cloud Run. The company named Anthropic, OpenAI, Modal, Snowflake, Ant Group and Tencent as external users running execution sandboxes or production deployments on the technology.

Moving gVisor into the CNCF could help push performance-related kernel optimisations upstream into Linux, according to Google. Those changes previously remained outside the mainline kernel because the project was controlled by a single vendor. Foundation governance removes that structural barrier.

The maintainer roster is also broadening. Google said Ant Group, Modal and Tines have committed to join gVisor's maintainers, while OpenAI, Tencent and Nvidia are continuing their contributions. For the CNCF, gVisor brings an eight-year-old project with named adoptees across the frontier-model and cloud markets.

What is OpenRL?

OpenRL addresses a different part of the AI stack: reinforcement-learning post-training, the stage where models are refined with reward-based feedback after initial training. Google described it as running Python-based reinforcement-learning post-training workflows on Kubernetes infrastructure, coordinating model sampling, outcome evaluation and weight updates.

The software uses Kubernetes Dynamic Resource Allocation to divide accelerator resources and schedule work across shared pools. Google said OpenRL supports standard engines including PyTorch and vLLM, with Google's Gemma and Alibaba's Qwen model families validated out of the box.

OpenRL also links with llm-d, a distributed inference and reinforcement learning project backed by Red Hat, Google Cloud, IBM Research, CoreWeave and Nvidia. Google said the pairing is intended to improve hardware use in two ways:

• routing rollouts to samplers with matching cache prefixes
• sharing GPUs between training and sampling phases

GPU efficiency has become a central cost question in AI development, since training and inference phases typically compete for the same expensive accelerators. Shared-pool scheduling directly targets that cost.

Why donate now?

OpenRL began in February as a research preview. Google said it is placing the project under CNCF governance early, before its design is fixed, so hardware makers, inference engine developers, training framework teams and model communities can help shape it.

That rationale differs from gVisor's. gVisor joins as a mature project seeking broader contribution and a path to mainline Linux. Agent Substrate arrives at version 1.0 with production availability on GKE. The three donations together give the CNCF a portfolio spanning agent execution, untrusted-code isolation and post-training orchestration.

The timing reflects market pressure. As companies deploy agents that write and execute code autonomously, the infrastructure decisions around density, isolation and governance are moving from engineering details to procurement-level concerns. Vendor-neutral governance signals to buyers that these building blocks will not be pulled back inside Google's walls.

Whether the Sandbox placements graduate toward incubating or graduated CNCF status will depend on adoption and maintainer diversity — the same path traveled by Kubernetes itself, which Google donated to the CNCF's predecessor in 2015 and which now underpins GKE, the platform hosting two of these three projects.