Google Cloud Launches Agent Compute Engine to Give Every AI Agent Its Own Isolated MicroVM

Quick Facts

• Google Cloud announced Agent Compute Engine (ACE) at the Linux Foundation's AGNTCon+MCPCon conference in San Jose, with a customer preview expected in the coming months.
• Each AI agent gets its own hardware-isolated microVM, enforced through Google's Titanium offload processor and Cloud Hypervisor, with sub-500ms suspend and resume latency.
• Google simultaneously released Agent Substrate, an open-source companion runtime that achieves up to 10x higher compute density than standard container runtimes.

Google Cloud introduced Agent Compute Engine at the Linux Foundation's AGNTCon+MCPCon conference in San Jose on Thursday, announcing a purpose-built infrastructure service designed to host AI agents at scale. The service is now available in preview and is built on the same technology that already runs Gemini for work tasks.

ACE assigns every AI agent its own microVM, a runtime environment that combines container speed with hardware-level security. Isolation is enforced through Google's Titanium offload processor, and each microVM runs on Cloud Hypervisor with an independent guest kernel and root privileges.

How It Works

The microVM architecture solves two core problems in agent infrastructure: security and resource waste. Malware spreading between sandboxes is blocked at the hardware boundary. Idle sandboxes, which traditionally consume significant compute even when inactive, are automatically shut down and reactivated in under 500 milliseconds when an agent resumes.

Security credentials are managed through a session-scoped SPIFFE certificate inside the guest. Long-lived user credentials never touch the host. Agent Gateway, which Google describes as an AI network firewall, swaps credentials into outbound requests after verifying the session's mTLS identity and egress policy.

Each microVM is also provisioned with its own Hyperdisk, giving agents persistent storage across long-running, multi-step sessions. Google built in cloning, snapshot and fork, and rewind and rollback capabilities so agent orchestration systems can operate efficiently without duplicating resources.

The Numbers

Google's official benchmarks for Agent Substrate, the open-source runtime that underpins ACE, show 10x higher compute density compared to traditional compute for AI agents. In a validation run of 200,000 agents, Agent Substrate sustained 500 activations per second at less than 500ms suspend and resume latency.

Google Kubernetes Engine also received updates tied to this launch. New agent sandboxes in GKE can deploy 300 sandboxes per second, per cluster, with sub-second time to first instruction.

"Agent Compute Engine microVMs give us the responsiveness, hardware isolation, and workload density needed to deliver interactive agent sandboxes to our users around the world," said Gautham Thambidorai, VP and Google Fellow for Gemini.

Open-Source Companion

Agent Substrate is the open-source sibling to ACE. It can scale to millions of sandboxes on a single Kubernetes cluster and supports agent frameworks including Claude Code, OpenClaw, and Hermes. Agent Substrate is self-managed; ACE is the managed cloud service. The two are built on the same underlying technology but serve different deployment preferences.

Governance Controls

Alongside ACE, Google announced expanded agent governance tools. Administrators can set monthly token spending caps per project and receive automated alerts at 50%, 80%, and 100% of the limit. Agent API calls halt automatically when the cap is reached. Governance spans Agent Runtime, Cloud Run, Agent Sandboxes in the Gemini platform, GKE, and ACE.

Upcoming previews include token spending controls, model routing, and oversight of coding agents on managed devices. The governance model is built on cryptographically attested agent identity, OAuth-propagated permissions, and action-level audit trails attributed to each agent.

Google Cloud CEO Thomas Kurian told reporters: "Now we're seeing as the models evolve people wanting to delegate tasks and sequences of tasks to agents, and these agents then being able to turn around and use a computer, use all of GCP and Workspace as a tool."

The announcements were part of a broader push Google Cloud has made around agentic infrastructure heading into 2027. A customer preview of ACE is expected in the coming months, with no exact date set.