# Google donates Agent Substrate and gVisor to CNCF for AI agent sandboxing

> **Key takeaway:** Google is moving Agent Substrate and gVisor to the CNCF to provide a scalable, secure orchestration layer specifically designed for the intermittent and untrusted workloads of AI agents.

**Published:** 2026-11-05T17:00:00+00:00 
**Source:** Google Open Source Engineering 
**Category:** bigtech-foss 
**Canonical URL:** https://fosswire.org/news/google-donates-agent-substrate-and-gvisor-to-cncf-for-ai-age.html 

## Summary
Google has announced the donation of Agent Substrate and gVisor to the Cloud Native Computing Foundation (CNCF) under Sandbox status. This move aims to address the unique infrastructure challenges posed by AI agents, which require high-density sandboxing and the ability to handle long periods of idle time without losing execution state.

Agent Substrate functions as an actor-based orchestrator built on top of Kubernetes, designed to manage millions of agent sessions. By decoupling stateful 'Actors' from compute-providing 'Workers,' the system allows agents to suspend and resume in under 500 milliseconds, avoiding the high costs of idle Pods and the latency of traditional serverless scaling.

## Technical architecture of Agent Substrate and gVisor
- **Actor-Worker Decoupling:** The system separates stateful Actors (containing filesystems, memory, and network rules) from Workers (pre-warmed Kubernetes Pods), allowing for rapid state restoration without constant Pod creation.
- **Sub-second State Restoration:** Using tiered storage, the Substrate controller can snapshot an Actor's memory and disk state upon idle and restore it to a warm Worker in less than 500 milliseconds.
- **Kernel Isolation via gVisor:** The platform utilizes gVisor (runsc) to intercept system calls in user space, providing a security boundary that prevents untrusted, model-generated code from compromising the host kernel.
- **Control Plane Optimization:** By managing agent lifecycles through an actor-based data plane rather than standard Kubernetes Pod objects, the system prevents etcd and kube-apiserver overload during massive reinforcement learning or evaluation loops.
- **Hybrid Virtualization Support:** Beyond gVisor, the architecture supports Cloud Hypervisor microVMs for workloads requiring hardware virtualization or specific kernel modules.

## Why it matters
This donation provides the cloud-native community with a standardized way to deploy untrusted AI agents at scale without the massive resource waste of traditional Kubernetes Pods. It offers developers a path to build highly responsive, secure agentic applications that can scale to millions of users while maintaining strict security isolation.