Agentic AI
Google Contributes Agent Substrate and gVisor to CNCF, Advancing Secure AI Agent Infrastructure
Google has handed over two critical technologies to the Cloud Native Computing Foundation to address the unique security and resource demands of running autonomous AI agents at scale.
InfoQ
Google has handed over Agent Substrate and gVisor to the Cloud Native Computing Foundation (CNCF), along with its OpenRL reinforcement learning project. The donations place these foundational tools under open-source stewardship, addressing two critical challenges in autonomous agent deployment: safely running code generated by AI models and efficiently managing the compute cycles that agents consume during idle periods.
Why AI Agents Need Their Own Infrastructure#
Autonomous agents operate fundamentally differently from traditional applications. Where conventional workloads run continuously and handle incoming requests, AI agents typically execute a task for seconds, then pause to await model inference, API responses, or human input. Google's analysis suggests agents spend 90–95% of their time idle yet must retain active state—memory, files, sessions—throughout.
This mismatch creates a resource problem. Dedicating a full Kubernetes Pod to each agent wastes capacity; scaling to zero introduces lag when environments must be rebuilt. Conversely, large-scale agent evaluations can strain Kubernetes control planes when thousands of short-lived execution environments churn in and out.
Agent Substrate: Decoupling State from Compute#
Google's Agent Substrate solves this by separating agent identity from execution resources. The system uses actors to maintain individual agent sessions with their state, and workers to provide reusable, pre-warmed execution pools drawn from shared Kubernetes infrastructure.
When an agent resumes work, Substrate assigns it to an idle worker and restores its prior state. When idle, the agent's state saves and the underlying compute resources release. This approach enables agents to resume in under 500 milliseconds without spawning new Pods, while reducing pressure on the Kubernetes control plane.
gVisor: Sandboxing Untrusted Code#
The second challenge is security. AI agents routinely execute shell commands, install packages, and call external tools—all code that may be generated by models and unvetted. Running such code as trusted presents major risk, especially when infrastructure is shared.
addresses this through a user-space application kernel that intercepts system calls instead of letting containerized agents interact directly with the host Linux kernel. This isolation boundary reduces host exposure while preserving the efficiency of container execution. Agent Substrate also supports for workloads requiring hardware-level virtualization, giving teams flexibility in their isolation strategy.

