Google Cloud launches Agent Substrate, an open-source GKE runtime for running millions of AI agent sandboxes

Google Cloud released Agent Substrate on September 15, 2026, an open-source execution runtime for GKE built to run millions of AI agent sandboxes at far higher density than standard containers.

What's new

Google's announcement describes Agent Substrate as "an open-source, secure-by-default agent execution runtime engineered to run millions of sandboxes with 10x higher density than standard container runtimes." The system is "purpose-built for the era of autonomous agents" and "delivers sub-500ms resume operations at over 500 suspend/resume activations per second with native zero-trust kernel and network isolation." In practice, that density lets a single host pack more than 1,000 dormant agents at once, waking any one of them in under half a second when it has work to do.

Architecturally, Google says "Agent Substrate avoids this by decoupling agent execution from machine management. Built on top of cloud-native Kubernetes infrastructure, Agent Substrate offers a new execution layer that's purpose-built for agentic workloads" — rather than treating each agent sandbox as a conventional long-running container that reserves a fixed amount of compute. For workloads that need to persist state between agent runs, Agent Substrate can be paired with Filestore agent volumes, a companion offering Google launched the same day that attaches and detaches NFS-backed storage volumes in milliseconds.

The runtime is open source, hosted at ate.dev, and runs on any standard Kubernetes cluster, though Google is optimizing it specifically for GKE. "Agent Substrate is open source and available to all GKE customers for non-production workloads. GA support for production is available via allowlist," meaning production use currently requires Google's sign-off rather than being self-service.

Context

The release follows a broader push across Google Cloud's September announcements toward agent-specific infrastructure — including BigQuery's "agent-ready" analytics and new billing controls for agent workloads — as cloud providers compete to make their platforms the default place to run large numbers of autonomous AI agents rather than a handful of long-lived services. Traditional container runtimes were designed around the assumption that most workloads run continuously; agent workloads instead spend much of their time idle between tasks, which is the gap Agent Substrate's suspend/resume model is built to close.

Why it matters

The specific numbers Google is publishing — 10x density, sub-500ms resume, 500+ activations per second — are a direct pitch to any team running large fleets of AI agents where most agents are idle most of the time: instead of paying to keep thousands of containers warm, Agent Substrate is designed to let a much smaller pool of hardware serve them by suspending and resuming sandboxes fast enough that the latency is barely noticeable. Making the runtime open source and portable to any Kubernetes cluster, rather than a GKE-exclusive feature, also lowers the switching cost for teams evaluating it, even as Google keeps production-grade GA access behind an allowlist while it manages the rollout.