Google donates gVisor, used by Anthropic and OpenAI, to CNCF

Google is handing three of its open source agent projects to the Cloud Native Computing Foundation (CNCF): the gVisor sandbox, the Agent Substrate runtime and the OpenRL training loop. The announcement came in a November 5 post from Google Cloud, which also declared Agent Substrate 1.0 generally available on GKE. Substrate 1.0 is usable now, either as a managed GKE capability or as open source on any Kubernetes cluster, but agents that need a GPU inside the sandbox can’t use it yet, because Substrate does not pass devices through to agents. gVisor’s promised “near native performance out of the box” depends on host kernel patches that upstream Linux has not merged.
Drew Bradstock, Google’s VP of product management for Kubernetes and orchestration, wrote the post, per The New Stack, and he pitches the donation as a replay of Kubernetes. “We believe that these donations will accelerate the development and adoption of agentic workloads through these projects similarly to what we saw with Kubernetes,” he writes. All three projects are now going through CNCF Sandbox onboarding, the foundation’s entry tier, with community maintainers and independent build infrastructure. Google says they were accepted last month, in October. The New Stack puts gVisor’s acceptance in late September. gVisor is on a path toward Incubation, the next tier.
How Agent Substrate runs agents
The problem Substrate addresses is idle time. An agent runs code for a few seconds and then waits on a model turn, an API or a human. In plain Kubernetes, a Pod per agent holds memory and CPU through those pauses, and the control plane has to track millions of persistent objects. Substrate splits the agent from the machine running it. Actors hold each agent’s memory, files and identity. Workers are reusable sandboxes taken from a warm Kubernetes pool. Gateways enforce inbound and outbound access and inject secrets based on the Actor’s identity, so keys never sit in the agent’s environment. When an agent pauses, Substrate snapshots it to storage and returns the Worker to the pool. When the next request arrives, it restores the agent onto any free Worker in under 500 milliseconds, Google says.
Google also claims 10x the density of standard container runtimes, production workloads across millions of sandboxes, and more than 500 suspend and resume activations per second. For isolation, teams choose Cloud Hypervisor microVMs or gVisor without changing how agent lifecycles are managed. Substrate runs standard OCI containers, so Google says ADK agents, stateful coding environments and productivity agents work out of the box, though each one still runs inside one of those two sandboxes. Nous Research is building Hermes Enterprise on it, the CNCF project kagent already uses it, and SPIRITT is building its agent workspace on the GKE version. SPIRITT CEO Tamir Magen credits it with “the density to run every customer’s agents economically at scale.”
GPUs are the gap. In the same post, Google says gVisor supports GPUs, TPUs and checkpoint/restore natively, and that Northflank uses gVisor to sandbox GPU workloads where nested virtualization isn’t available. Substrate’s API guide, however, marks devices as “temporarily unsupported”, so the accelerator support Google sells for gVisor doesn’t yet reach agents running under Substrate.
Why gVisor’s kernel patches matter
gVisor is an application kernel written in Go. It sits between a container and the host and handles the container’s system calls in userspace, so untrusted code never talks to the host kernel directly. Ordinary containers share one host kernel, so a single kernel vulnerability can be enough for an escape. Unlike a VM, gVisor doesn’t preallocate a fixed block of RAM, and it needs no nested virtualization. Google open-sourced it in 2018, and it runs millions of sandboxes a day under GKE Sandbox and Cloud Run. Outside Google, Anthropic and OpenAI use it to contain code execution, Modal and Snowflake run their execution sandboxes on it, and Ant Group and Tencent run it in production. The New Stack reports that Anthropic runs code generated by Claude inside gVisor containers on its web platform. It also reports that an OpenAI security report published on November 4 cited sandboxes such as gVisor and Firecracker as important for limiting the damage from exploited Linux kernel vulnerabilities, and that in July models in an OpenAI cybersecurity evaluation escaped their restricted environment and breached Hugging Face’s production infrastructure.
Speed has been the cost. InfoQ noted at launch in 2018 that gVisor may perform poorly on workloads heavy in system calls, and that some calls are unsupported, so not every application runs inside it. Google says that sandbox providers, Google among them, carry host kernel patches that bring gVisor’s I/O close to native Linux. According to Google, upstream maintainers declined to merge them while a single vendor owned the project, and neutral governance opens a path to upstreaming them. Ant Group, Modal and Tines have committed to join gVisor’s maintainers, and OpenAI, Tencent and NVIDIA will keep contributing. Under the plan gVisor’s contributors published in October, non-Google maintainers get merge permissions over the next few months, CI moves to GitHub Actions and Buildkite, and Google’s internal tests no longer block pull requests.
OpenRL, the early one
OpenRL runs self-hosted reinforcement learning post-training loops in plain Python on Kubernetes. Its four primitives (sample, forward_backward, optim_step, save_state) sit behind a Tinker-compatible API, so existing tinker-cookbook recipes run unchanged, with LoRA adapters or full fine-tuning. Training uses PyTorch/AutoModel and sampling uses vLLM. Gemma and Qwen are validated, and TPU support is on the roadmap. Paired with llm-d, it routes multiturn rollouts to samplers that already hold matching KV cache prefixes, and it time-slices the same GPUs between sampling and training. Google calls it the newest layer of the stack “by a wide margin” and says it is bringing OpenRL to the CNCF before the design hardens.
Our analysis
What we checked
On November 7 we read gVisor’s CNCF application and onboarding issues and its maintainers file, LWN’s 2022 report on a gVisor kernel patch, mainline Linux’s KVM hypercall header, and the GitHub repositories and docs of Agent Substrate and OpenRL.
Google presents the stalled kernel patches as a governance problem, but that is its own account. The one public gVisor kernel proposal we found was rejected on security and design grounds, and Google’s CNCF application calls full performance parity “structurally impossible and not a goal”. The maintainer change hasn’t happened yet either. Substrate’s contributor count and dates check out, but its restore figures have no published run behind them.
- Andrei Vagin’s July 2022 patch series let gVisor issue host system calls through KVM. LWN reported Thomas Gleixner’s reply as “Seriously no”, and Sean Christopherson called the approach “mildly terrifying”. Mainline defines 12 KVM hypercalls, and none of them is a host-syscall one. This is one series, and Google may hold other patches it never posted.
- The application was filed on September 7 with Ant Group, Tines and Modal as co-sponsors. It says the patches were refused because of “gVisor being perceived as a Google-only project”. It also puts gVisor’s performance problem at “75% perception/25% reality”.
- gVisor’s maintainers file lists 18 active maintainers, all at Google, and the repository is still at
google/gvisor. A neutralgvisorGitHub org had DCO sign-off checks enabled on November 6, but moving the project there is an open onboarding task. - The CNCF vote ran from September 22 to 28. It passed with 72.73% of binding voters in favour and none against, from 8 binding votes.
- Agent Substrate’s repository was created on May 13 and has 131 contributors. v1.0.0 was published at 23:55 UTC on November 5, after three release candidates that began October 29. The first tagged release, v0.1.0, came on September 10. It has 16 maintainers from Google and 3 from Solo.io. The repo includes a benchmark harness that records restore percentiles, but the only saved runs are for a tunnel benchmark, so the sub-500 ms restore figure is documented, not measured.
- Google says OpenRL “started in February as a research preview”. Its first commit is dated February 10, and the public research preview came on June 11. It has 10 contributors and is at v0.0.8, released November 4.
Up next
Scientists use Claude least where they say it matters most
A researcher opens Claude Code to crunch data, like nearly half of the 78,135 science sessions Anthropic studied. But ideation, which scientists rank most important, gets 2.1%. Roughly what share of your Claude sessions goes to brainstorming ideas?
Related stories
- Google opens Gemini’s agent sandboxes to Cloud customers
- OpenAI gives its Open Responses API spec to the AAIF
- OpenAI’s dots can now read email codes to pass 2FA
- Gemini 4 Carbon tops AA Index, goes to cyber defenders first
- Claude Code’s sandbox ships as an open-source npm package
- Claude Duet won’t share your card number even if you ask
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

