Google Cloud Launches Agent Compute Engine: A Dedicated MicroVM for Every AI Agent

By Michal Sutter | October 22, 2026

Google Cloud has unveiled Agent Compute Engine, a compute service purpose-built for AI agents. The core proposition is simple but significant: every agent session runs in its own hardware-isolated microVM, complete with a separate guest kernel, root access, and Hyperdisk storage. The sandbox carries zero credentials, and the service already powers Gemini as an agent for work tasks. Google announced the offering this week at AGNTCon + MCPCon, with customer preview targeted for the coming months. The headline promise: VM-grade isolation with container-like speed and density.

TL;DR

• Unit of isolation: One hardware-isolated microVM per agent session, each with its own guest kernel, root access, and Hyperdisk.
• Runs on: Google-managed hosts powered by Titan silicon and the Titanium IPU. An open-source sibling, Agent Substrate, runs on any Kubernetes cluster.
• Performance: Idle agents are checkpointed to zero active resources and resumed from snapshots. Google has not yet disclosed exact latency figures.

Why a MicroVM per Agent?

Agentic workloads introduce a challenge that traditional container models handle poorly. Agents execute untrusted code, call external tools, and accumulate state across long-running sessions. Isolating each agent in a dedicated microVM gives operators VM-level security boundaries without sacrificing the fast startup and high density that container-based deployments typically provide. The credential-free sandbox design further reduces the blast radius if an agent is compromised or behaves unexpectedly.

How It Works

Agent Compute Engine runs on Google-managed infrastructure using Titan silicon and the Titanium IPU, Google's custom networking and security offload stack. Each session receives its own microVM with a guest kernel and root access, backed by Hyperdisk storage. When an agent goes idle, the platform checkpoints it to zero active resources and later resumes it from a snapshot—an approach that keeps costs down for sporadic or long-lived agents.

For organizations running their own infrastructure, Google is also offering Agent Substrate, an open-source counterpart that runs on any Kubernetes cluster, bringing the same isolation model to multi-cloud and on-premises environments.

Availability

Agent Compute Engine is in customer preview, with general availability expected in the coming months. Google's positioning is clear: as agentic AI moves from experimentation to production, the underlying compute layer needs to be rethought around the agent as the primary unit of work—not the container or the virtual machine of the previous era.

via MarkTechPost