#

Key takeaways

• Google Cloud is donating three projects to the Cloud Native Computing Foundation.

• The projects are Agent Substrate, gVisor, and OpenRL.

• Google says the projects are entering Sandbox onboarding with community maintainers and independent build infrastructure.

• Agent Substrate reached version 1.0 and is generally available as a managed offering on Google Kubernetes Engine.

• Google says gVisor is on a path toward CNCF Incubation.

Google Cloud moves three projects into CNCF

Google Cloud announced that it is donating three projects to the Cloud Native Computing Foundation. The projects are Agent Substrate, gVisor, and OpenRL. Google says the projects were accepted into the CNCF Sandbox the previous month, and they are now going through Sandbox onboarding.

The company says the onboarding process includes community maintainers and independent build infrastructure. It also says gVisor is on a path toward CNCF Incubation. The announcement separates the CNCF governance news from the product availability news for Agent Substrate.

What Agent Substrate does

Google says Agent Substrate addresses the cost of leaving an agent environment running while it waits for a model response, API call, or person. The project separates persistent session state and identity from active compute. That design lets the system snapshot idle sessions to storage and restore them when needed.

Google reports restores in under 500 milliseconds. It also describes higher sandbox density as a benefit of the architecture. Those figures and claims come from the announcement, and they describe the project's intended operating model.

Google separately announced that Agent Substrate reached version 1.0. It is generally available as a managed offering on Google Kubernetes Engine. The announcement ties that availability to GKE, while the CNCF news covers the project's governance path.

gVisor and OpenRL in the same stack

Google describes gVisor as a way to isolate untrusted code in a lightweight userspace application kernel on Linux. That places the project in the security and isolation part of the stack. Google says gVisor handles application system calls in userspace, keeping untrusted code from communicating directly with the host kernel.

OpenRL coordinates reinforcement learning after training. Google says it supports self-hosted post-training loops in Python on Kubernetes and parallel sandboxed rollouts. In the announcement, OpenRL sits alongside Agent Substrate and gVisor as part of a broader open stack.

A stack for Kubernetes and multiple accelerators

Google describes the open projects as a stack designed to operate across Kubernetes clusters and clouds. It also says the stack is designed to work across CPUs, GPUs, and TPUs. That framing links the three projects into one infrastructure story rather than treating them as separate releases.

The announcement does not make a broader market claim. It focuses on the projects, their CNCF onboarding, and the availability of Agent Substrate on GKE. The supported facts show two parallel tracks: open governance for all three projects and managed availability for one of them.

Operational considerations from the announcement

The main operational point in the source is the separation of persistent state from active compute. That design can reduce the need to keep an agent environment running while it waits. Google also highlights snapshot and restore behavior, which suggests a workflow built around idle session storage and later recovery.

The source also points to governance and build-process changes. Community maintainers and independent build infrastructure are part of the CNCF onboarding path. For readers tracking project maturity, the announcement presents both technical architecture and open-source stewardship as part of the same update.

What to watch next

The source gives one clear next step for gVisor: a path toward CNCF Incubation. It also shows that Agent Substrate has already reached version 1.0 and general availability on GKE. OpenRL is included in the CNCF donation and onboarding process, with its role defined by post-training reinforcement learning workflows.

Taken together, the announcement shows Google Cloud aligning these projects under CNCF governance while also shipping one of them as a managed service. The facts in the source stay focused on the projects, their roles, and their current status.