Google Gives Agent Substrate, gVisor and OpenRL to the CNCF
Google is donating Agent Substrate, the gVisor sandbox and its OpenRL reinforcement learning project to the Cloud Native Computing Foundation. The move targets secure, long-running execution of AI agents on Kubernetes.
TL;DR
- Google is donating Agent Substrate, gVisor and OpenRL to the Cloud Native Computing Foundation (CNCF).
- The projects target two problems: isolating untrusted, AI-generated code and managing agents that spend most of their time idle.
- Teams running agents on Kubernetes should expect new isolation and lifecycle patterns to appear in the cloud-native ecosystem.
Why agents strain Kubernetes
Kubernetes was built for applications that start, serve requests and keep running. AI agents behave differently. As InfoQ describes it, an agent may run a tool or some code for a few seconds, then wait for model inference, an external API or human approval. Google estimates that agents can spend 90-95% of their time idle, while still holding state such as memory, files, repositories and active sessions.
That pattern is awkward for the platform. Giving each agent its own Pod wastes resources. Scaling to zero saves compute but adds delay, because the environment has to be rebuilt when work resumes. Large agent evaluations create the opposite problem: thousands of short-lived environments that are created and destroyed quickly can put heavy load on the Kubernetes control plane.
Separating agent state from compute
Agent Substrate splits the problem in two. Actors hold an agent session, including filesystem contents, memory and execution context. Workers are reusable, pre-warmed environments drawn from shared Kubernetes pools. When an agent resumes, it is assigned to an available worker and its previous state is restored. When it goes idle, the state can be saved and the compute released.
Google says this lets agents resume in under 500 milliseconds without creating a new Pod for each execution. Because lifecycles are managed apart from individual Pod objects, the approach also eases pressure on the control plane.
Isolating generated code with gVisor
Agents often run generated shell commands, install packages and call external tools. Treating that code as trusted is risky, especially when several agents share the same infrastructure.
gVisor is a user-space application kernel. Instead of letting containerised applications call the host Linux kernel directly, it intercepts system calls and handles them in its own isolated environment. That reduces exposure of the host while keeping much of the efficiency of container-based execution. Agent Substrate can also use Cloud Hypervisor microVMs when hardware-level virtualisation is required, so platform teams can pick the isolation boundary without changing the agent lifecycle model.
Where this fits in the ecosystem
Other providers address similar problems with Kata Containers, Firecracker and dedicated sandboxing platforms. InfoQ's reading is that Google's contribution is the combination: secure execution paired with an agent-aware orchestration model, placed under open, vendor-neutral governance at the CNCF. OpenRL, a reinforcement learning orchestration project, is part of the same announcement and targets coordination of large training workloads.
The donation is framed as an effort to give agentic applications an open foundation, much as Kubernetes standardised container orchestration. Whether that foundation gains broad adoption will depend on how the projects mature within the CNCF, a point the announcement does not settle.
Why it matters
For platform and backend teams, agents are starting to need their own execution layer: environments that can be suspended, restored and isolated far more often than typical services. If Agent Substrate and gVisor follow the usual CNCF path, teams running agent workloads on Kubernetes may gain a standard way to handle sandboxing and state without building custom pooling and lifecycle code. Teams should watch how these projects evolve, and review how their current agent runners isolate generated code, since shared infrastructure raises the stakes.
Sources
This article was written by Blozhik from the reporting below. Read the originals for full detail.
InGoogle Donates Agent Substrate and gVisor to the CNCF to Secure AI Agents at ScaleInfoQ