AI
Google Donates Agent Substrate and gVisor to CNCF
The move addresses resource management and security challenges as AI agents create fundamentally different infrastructure demands than traditional cloud workloads.
Key takeaways
- Google donated Agent Substrate and gVisor to CNCF to address AI agent infrastructure challenges under open governance
- Agent Substrate separates agent state from compute resources, enabling sub-500ms resume times without creating new Kubernetes Pods
- AI agents spend 90-95% of their time idle while holding state, creating resource management problems traditional Kubernetes wasn't designed to solve
- gVisor provides user-space kernel isolation for executing untrusted AI-generated code without exposing the host system
- The move reflects recognition that agentic systems require fundamentally different infrastructure than traditional cloud-native applications
Google moves AI agent infrastructure to open governance
Google has transferred Agent Substrate and gVisor to the Cloud Native Computing Foundation, placing critical components of its AI agent infrastructure under vendor-neutral governance. The donation reflects a recognition that autonomous agents create infrastructure demands that existing container orchestration systems weren't designed to handle.
The company also contributed OpenRL, its reinforcement learning orchestration project, as part of a broader effort to establish open standards for agentic applications similar to how Kubernetes standardized container management.
The idle agent problem
Traditional Kubernetes workloads follow predictable patterns: applications start, serve requests, and run continuously. AI agents behave differently. They execute code or tools briefly, then wait—for model inference, external API responses, or human approval. Google estimates agents spend 90-95% of their time idle while maintaining critical state including memory, files, and active sessions.
Keeping each agent in a dedicated Pod wastes resources. Scaling to zero introduces delays when environments must be rebuilt. Large-scale agent evaluations can overwhelm Kubernetes control planes when thousands of short-lived environments are created and destroyed repeatedly.
Agent Substrate addresses this by separating agent state (actors) from compute resources (workers). Actors represent individual agent sessions and retain their state. Workers provide reusable, pre-warmed execution environments from shared Kubernetes pools. When an agent needs to resume, Substrate assigns it to an available worker and restores its state. When idle, the state is saved and compute resources are released.
According to Google, this approach enables agents to resume in under 500 milliseconds without creating new Kubernetes Pods for each execution, while reducing control plane pressure.
Securing untrusted code execution
AI agents routinely execute generated shell commands, install packages, and interact with external tools. Treating this code as trusted creates significant security risks, particularly when multiple agents share infrastructure.
Google's gVisor provides an additional isolation boundary through a user-space application kernel. Instead of letting containerized applications interact directly with the host Linux kernel, gVisor intercepts system calls and handles them within its own isolated environment. This reduces host exposure while maintaining much of the efficiency associated with container-based execution.
Agent Substrate also supports Cloud Hypervisor microVMs for workloads requiring hardware-level virtualization, giving platform teams flexibility to select appropriate isolation boundaries without changing the agent lifecycle model.
Why it matters
The donation signals that AI agents require fundamentally different infrastructure than traditional applications. Kubernetes was designed for long-running, relatively stable workloads. Agentic systems demand environments that can be created, suspended, restored, and isolated at much higher frequency.
Other providers are addressing similar challenges through technologies like Kata Containers and Firecracker. Google's contribution combines secure execution with agent-aware orchestration and places both within an open ecosystem, potentially establishing foundations for the next layer of cloud-native computing.
By moving these projects to the CNCF, Google is acknowledging that AI agents need more than orchestration frameworks and tool access—they need secure, stateful, and efficiently managed execution infrastructure that the industry can build on collectively.
These details were first reported by InfoQ.