News
What Is Agent Compute Engine? Google’s MicroVM for Every AI Agent
Agent Compute Engine is Google Cloud’s new service that gives each AI agent its own hardware-isolated microVM. What it does, when it opens, and what it costs.
Google Cloud announced Agent Compute Engine on October 22, 2026, and the name sounds a lot like the Google Compute Engine that has rented out servers since 2013. So what is the new one, and why does an AI agent need its own computer at all?
The short version: an agent that writes code, runs commands and opens websites needs a safe place to do that. Agent Compute Engine (some headlines shorten it to ACE) is Google’s answer. Here is what it is, how it works, and what it means if you are not running a data center.
What is Agent Compute Engine?
Agent Compute Engine is a compute service built only for AI agents. Each agent gets a microVM: a tiny virtual computer that starts fast and is walled off from every other agent on the same physical server. Inside it, the agent has root access, so it can install tools, run scripts or drive a browser without being able to reach its neighbors.
Google says the service started as the internal system behind Gemini as an agent for work tasks. The announcement post was written by three Google Cloud infrastructure leaders: Nirav Mehta, Michael Bachman and Kevin Hutchins. It was shown at AGNTCon + MCPCon, the Linux Foundation’s agent conference in San Jose.
Think of it like a hotel. Older cloud setups were like a shared apartment, where one messy roommate could affect everyone. Agent Compute Engine gives each agent its own locked room, with its own key, and the hotel tidies the room away while the guest is out.
How does Agent Compute Engine work?
Google’s announcement lists a handful of pieces. Here they are in plain English:
| Feature | What Google says | Why it matters |
|---|---|---|
| MicroVM per agent | Each agent runs behind a hardware-isolated Cloud Hypervisor boundary with its own guest kernel | A broken or tricked agent cannot spill into other agents |
| Titanium chips | Runs on Titanium, Google’s offload processor, for "silicon-rooted" defense | Security starts in hardware, not just software |
| Pause and resume | Idle microVMs are checkpointed and resumed from snapshots when the agent is waiting on a person or a system | You do not pay to keep a waiting agent running |
| Shared memory | Copy-on-write memory lets many agents share one operating system template | More agents fit on each server, so costs drop |
| No secrets inside | Service credentials are added at the network exit, outside the microVM | An agent cannot leak a password it never sees |
| Own identity and disk | Each microVM gets short-lived credentials, a verified ID and its own Hyperdisk for saved work | Every action can be traced to one agent, and work survives a pause |
The "no secrets inside" idea is the same one NVIDIA uses in its open-source sandbox, which we covered in What is NVIDIA OpenShell? Security teams are settling on one rule: never hand the agent the real keys.
Agent Compute Engine vs Agent Substrate
Google announced a second, related tool the same day, and the two are easy to mix up.
- Agent Compute Engine is a managed Google Cloud service. Google runs the machines. It is in a waitlist stage, with preview "in the coming months."
- Agent Substrate is open-source software for Kubernetes (Google’s GKE, or any Kubernetes cluster). You run it yourself. Google says it works with agent tools such as Claude Code, OpenClaw and Hermes.
The headline numbers belong to Agent Substrate, not Agent Compute Engine. Google says Substrate reached up to 10x higher compute density than traditional setups, and in a test with 200,000 agents it held 500 activations per second with suspend and resume under 500 milliseconds. Those are Google’s own figures. Google has not published speed or density numbers for Agent Compute Engine itself.
What else Google announced: agent governance
The same post expanded Google’s agent controls for businesses. Agent Gateway acts as the "air traffic controller" for agent traffic, Agent Registry lists approved agents, tools, MCP servers and skills, and Agent Identity gives each agent an ID based on the SPIFFE open standard. New this quarter, in preview, are token budgets, model routing rules and sign-in partners Ping Identity, SailPoint and Saviynt. Google also says it will bring these controls to coding agents like Claude Code on company laptops.
The pattern matches other launches this week, such as 1Password for Agents: the industry is busy deciding who an agent is and what it is allowed to touch.
Long agent jobs are getting their own plumbing on the tool side too: MCP Tasks let an MCP server hand back a ticket and finish the work in the background.
Is Agent Compute Engine free? How do you try it?
No pricing has been announced. Google has only said a preview will open "in the coming months," and businesses can fill in an interest form linked from the announcement. There is no consumer version. If you are a developer who wants to test the idea today, the open-source Agent Substrate is the part you can actually run.
What this means for you
For most people, Agent Compute Engine is plumbing. You will never log in to it. But it tells you where AI is going. Google built a whole new kind of server because its Gemini agents now do long jobs that run code and use the web, and that work needs walls around it.
You do not need your own microVM to get that kind of help. What you need is an app that already picked the models, set up the agents and wrapped them in a simple screen. That is the idea behind agentic AI for regular people.
Agentic AI gives you a team of ready-made expert agents on your iPhone and iPad, running on frontier models, with no account and no login. You describe the task, the agents plan it and hand back finished work. Big companies get sandboxes and gateways; you get the result.
Quick answers
What is Google Agent Compute Engine? A Google Cloud service, announced October 22, 2026, that runs each AI agent in its own hardware-isolated microVM with its own kernel, disk and identity. It already powers Gemini’s agents.
Is Agent Compute Engine the same as Google Compute Engine? No. Google Compute Engine is Google’s general virtual machine service. Agent Compute Engine is a separate service made only for AI agents, with fast pause and resume and no secrets stored inside the machine.
When is Agent Compute Engine available? Google says a preview opens "in the coming months." There is no exact date yet. Businesses can register interest through Google’s form.
How much does Agent Compute Engine cost? Google has not announced pricing.
What is Google ACE for AI agents? "ACE" is a shorthand some outlets use for Agent Compute Engine. Google’s own posts use the full name.
Is Agent Compute Engine like NVIDIA OpenShell? Both keep credentials out of the agent’s reach and wall agents off. OpenShell is free, open-source software you run yourself; Agent Compute Engine is a managed Google Cloud service. Google’s open-source equivalent is Agent Substrate. More terms in our agentic AI glossary.
Sources: Google Cloud blog: Introducing Agent Compute Engine · Google Cloud: Agent Compute Engine deep dive · The Decoder · The New Stack · Crypto Briefing
