AI NewsInfrastructureAnnouncement

Google donates Agent Substrate, gVisor and OpenRL to the CNCF

Google hands three open-source agent infrastructure projects to the CNCF. Anthropic, OpenAI and Snowflake are named production users of gVisor.

AI News

Editorial2 min read

LinkedInX

Why it mattersA team running many agents gets a vendor-neutral sandbox and runtime. Anthropic, OpenAI and Snowflake already use gVisor in production.

A team running many AI agents on Kubernetes pays for a Pod that sits idle most of the time. Google wants to change that. It is donating Agent Substrate, a new runtime, to the Cloud Native Computing Foundation. The runtime keeps an agent's memory and files, but drops the compute when the agent waits on a model call.

The CNCF accepted three Google projects into Sandbox status last month. Google announced the donation ahead of KubeCon North America in Salt Lake City. The three projects are Agent Substrate, gVisor and OpenRL.

Agent Substrate separates the stateful part of an agent from the worker that runs each step. The memory, files and identity stay. The worker is a shared pool. Google says the runtime restores an idle agent in under 500 milliseconds. It performs more than 500 suspend and resume operations per second. It reaches ten times the density of standard container runtimes.

gVisor goes to a neutral owner

gVisor is Google's user-space kernel for running untrusted code. Google's blog names the production users. Anthropic says it runs Claude's code execution inside gVisor. OpenAI secures research infrastructure with it. Modal uses it for its execution sandboxes. Snowflake runs its Snowpark sandbox on it. Ant Group and Tencent are named as long-term maintainers. NVIDIA, OpenAI and Tines also contribute.

The move removes a long-standing block on kernel work. gVisor's performance depends on Linux kernel patches that bring its I/O close to native speed. Upstream Linux maintainers have declined to merge those patches while the project belonged to one company. Moving to CNCF governance opens the path to upstreaming them.

OpenRL closes the set

OpenRL coordinates reinforcement learning loops and evaluation sandboxes on Kubernetes. The three projects together cover parts of running agents at scale that Kubernetes was not designed for. Agents wake up, call a model, and wait. Google says AI agents spend 90 to 95 percent of their time waiting. They wait on a tool call, a model turn or a human. Agent Substrate aims to make that idle time cheap.

Richard Macmanus at The New Stack writes that Google wants to repeat the Kubernetes playbook. Donate the project. Let a neutral foundation maintain it. Grow the ecosystem. Google's own blog does not make that claim in those words. But the three projects together read as a stack for agent workloads with no single vendor behind them.

A team planning its agent runtime has a reason to watch this. Agent Substrate, gVisor and OpenRL are all Apache 2.0. The CNCF Sandbox tier is the first of three. The projects still have to graduate through Incubating to Graduated. The maintainers list names the companies already paying the cost of running agents at scale. That is the useful signal. Anthropic and OpenAI use gVisor today to keep agent code off the host kernel. Those tools now have a neutral owner.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project