Google Cloud donates three AI projects to CNCF Sandbox

Google has donated Agent Substrate, gVisor and OpenRL to the Cloud Native Computing Foundation. All three open-source projects have been accepted into the CNCF Sandbox.

The tech giant said the move places three projects focused on running and training AI agents under vendor-neutral governance at a time when companies are managing growing numbers of automated software tasks on shared infrastructure.

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, the executive said.

Agent Substrate 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. Agent Substrate has also reached version 1.0, generally available on GKE.

The software runs standard OCI containers and can use either Cloud Hypervisor microVMs or gVisor as the isolation layer. The managed GKE service targets production deployments, while the open-source code can run on any Kubernetes cluster.

Managed and open

Bradstock argued that agent workloads differ from conventional cloud applications because they execute code in short bursts, pause frequently, and often need strict controls over tools, secrets and external services.

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 matters for AI agent workloads because agents may generate shell commands or run unreviewed third-party packages, said Bradstock. 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. It also 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 have previously remained outside the mainline kernel because the project was controlled by a single vendor.

Google added that Ant Group, Modal and Tines have committed to join gVisor's maintainers, while OpenAI, Tencent and Nvidia are continuing their contributions.

Training focus

OpenRL addresses another part of the AI stack: reinforcement learning post-training. Google described it as running Python-based reinforcement-learning post-training workflows on its own 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 Gemma and Qwen validated out of the box.

It 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 by routing rollouts to samplers with matching cache prefixes and by sharing GPUs between training and sampling phases.

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