Google Donates Agent Substrate, gVisor, and OpenRL to CNCF

Google has donated Agent Substrate, gVisor, and OpenRL to the Cloud Native Computing Foundation (CNCF) after the projects were accepted into the CNCF Sandbox last month. Agent Substrate has also reached 1.0 and is generally available as Agent Substrate on GKE, part of an open stack that runs on any Kubernetes cluster, on any cloud, across CPUs, GPUs, or TPUs. The rationale is that agent workloads change the platform underneath: an agent works in a continuous, stateful loop, proposing an action, spinning up subagents, executing code or tool calls inside a sandbox, and feeding results back into the model. Running that loop at scale, the announcement says, requires the same open foundation and neutral governance that Kubernetes and the CNCF brought to containers.

Agent Substrate decouples stateful agents from active compute. Actors hold each agent’s memory, files, and discrete identity; Workers are reusable execution sandboxes drawn from a warm Kubernetes pool; Gateways enforce inbound and outbound access control and inject secrets based on Actor identity. When an agent pauses to wait on a model turn or tool call, Agent Substrate snapshots the agent’s memory and files to storage and returns the Worker to the warm pool. On the next request it restores the Actor onto any available Worker in under 500 milliseconds. Since public development started in May 2026, the project has grown to more than 100 contributors; Nous Research is building Hermes Enterprise on it, and it already powers the CNCF project kagent. The 1.0 milestone is described as accepting production workloads across millions of sandboxes at 10x the density of standard container runtimes, with more than 500 suspend and resume activations per second. Users can choose Cloud Hypervisor microVMs or gVisor as the isolation boundary, and because Agent Substrate runs standard OCI containers, it works with ADK agents and stateful coding or productivity agents. It is now available as a managed capability on GKE, complementing Agent Compute Engine, Google’s Titanium microVM service. SPIRITT is building its SPIRITT Workspace on Agent Substrate on GKE; CEO Tamir Magen is quoted on the security isolation, extensibility, and density it provides.

gVisor addresses the risk of agents running generated shell commands and unreviewed packages. Standard containers share the host Linux kernel with other workloads on the node, which the post calls a risky bet for untrusted code. gVisor is a lightweight application kernel written in Go that sits between the container and the host OS, intercepting and handling application system calls in userspace. Open sourced by Google in 2018, it runs millions of sandboxes daily under GKE Sandbox and Cloud Run. Anthropic and OpenAI use it to contain code execution; Modal and Snowflake run execution sandboxes on it; Ant Group and Tencent run it in production. The post highlights three capabilities: container density with flexible memory, native GPU, TPU, and snapshot support, and deployment on any Linux node without nested virtualization. Northflank is cited for using gVisor to sandbox GPU workloads when nested virtualization is unavailable. Donation to the CNCF matters because host kernel patches that bring gVisor‘s I/O performance close to native Linux were not merged while the project was owned by a single vendor; neutral governance opens the path to upstreaming those optimizations. Ant Group, Modal, and Tines have committed to join the maintainers, with OpenAI, Tencent, and NVIDIA continuing contributions.

OpenRL provides a self-hosted Python reinforcement-learning post-training loop on Kubernetes. The post frames RL post-training as the source of many frontier capability gains: running millions of multiturn rollouts inside real tools, scoring outcomes, and updating model weights. OpenRL exposes four primitives — sample, forward_backward, optim_step, and save_state — through a Tinker-compatible API, keeping weights and data on the user’s cluster and running existing tinker-cookbook recipes unchanged for LoRA or full parameter fine-tuning. It uses Kubernetes Dynamic Resource Allocation to partition accelerator pools, with PyTorch/AutoModel for training and vLLM for sampling; Gemma and Qwen are validated out of the box, with TPU support on the roadmap. OpenRL pairs with llm-d, the CNCF distributed inference and RL project founded by Red Hat, Google Cloud, IBM Research, CoreWeave, and NVIDIA. llm-d contributes prefix-aware rollout routing, so samplers holding matching KV cache prefixes handle multiturn rollouts, and GPU time-slicing so the same physical GPUs can alternate between sampling and training. OpenRL is the newest piece: it started in February as a research preview and is being brought to CNCF early to avoid a design hardened by a single company, with a north star of one recipe for any model on any accelerator.

The post explains why the CNCF is the right home now: improving agent security and efficiency requires joint design across runtime, scheduler, compiler, and silicon, and vendor-neutral ownership lets cloud providers, hardware vendors, and independent software companies participate equally. Each project is working through CNCF Sandbox onboarding with community maintainers and independent build infrastructure, and gVisor is on a path toward Incubation. The article closes with KubeCon + CloudNativeCon North America in Salt Lake City from November 9 to 12, and pointers to the CNCF Sandbox applications (gVisor #521, Agent Substrate #523, OpenRL #518, and KubeRay #525), the project sites ate.dev, gvisor.dev, and openrl.dev, and the managed offerings Agent Substrate on GKE and GKE Sandbox.

Open infrastructure for autonomous agents comes to CNCF

View Original