Powering open infrastructure for autonomous agents

AI agents operate in continuous loops, requiring scalable infrastructure that differs from traditional systems. Google is addressing this by donating Agent Substrate, gVisor, and OpenRL to the CNCF, following their acceptance into the Sandbox. These projects aim to provide the foundational tools needed for autonomous agent workloads, similar to how Kubernetes and CNCF transformed container management. The donation reflects a shift toward open governance and shared infrastructure for agentic systems.

Agent Substrate, now at version 1.0 and generally available on Google Kubernetes Engine (GKE), decouples stateful agent sessions from active compute. It snapshots idle agents to storage and restores them in under 500 milliseconds, enabling 10x higher container density. This approach avoids the inefficiency of keeping dedicated containers alive during pauses, which ties up memory and compute resources in standard Kubernetes setups.

gVisor isolates untrusted code within a lightweight userspace application kernel, reducing security risks associated with running unreviewed code in containers. It supports container packing density on any Linux node, including GPUs and TPUs, without requiring nested virtualization. The project, open-sourced by Google in 2018, is used by major companies like Anthropic, OpenAI, and Tencent for secure sandboxing of agent workloads.

OpenRL provides an open-source framework for reinforcement learning post-training loops in plain Python on Kubernetes. It orchestrates self-hosted RL pipelines, enabling agents to learn across millions of parallel sandboxed rollouts. The framework supports efficient distribution of training and serving workloads, complementing Agent Substrate and gVisor in building scalable, secure agentic systems.