Google Agent Compute Engine: One microVM per Agent, No Credentials Inside
Direct answer
Google Cloud announced on October 22, 2026, at AGNTCon + MCPCon (San Jose), Agent Compute Engine: a compute service built to run AI agents, where each agent gets its own hardware-isolated microVM. The service already powers Gemini as a work agent and Antigravity, and will open its customer preview in the coming months (no exact date). The thesis: agents are no longer just another script on your server; they need their own computer, their own identity, and no secrets inside for an attack to steal.
How it works, in four pieces
- A computer per agent: each microVM has its own guest kernel and root access, runs on Cloud Hypervisor with Google's Titanium hardware, and can run scripts, drive a headless browser and write code in loops that can last weeks.
- No credentials inside the sandbox: the sandbox never holds user credentials or API keys. Outbound requests go through a local proxy to the Agent Gateway off-node, which validates user delegation, agent identity and node identity, then inserts the backend credential in flight (token swap).
- Pause and resume by snapshot: when the agent finishes a tool call and enters a waiting phase (thinking or waiting on a person), the microVM is paused by checkpoint and resumed from snapshot. That matters for cost: idle agents should not bill like machines left on all the time.
- Density through shared memory: with copy-on-write memory sharing, many agents starting from the same base share memory until one changes something. Google reports up to 10x higher compute density with this design, a vendor claim without published independent validation.
The other half of the announcement: governance wherever your agents run
Alongside Agent Compute Engine, Google extended its Gemini for business governance stack (Agent Gateway, Agent Registry, Agent Identity and observability) to cover agents running elsewhere: Cloud Run and Google Kubernetes Engine, with cross-cloud, on-prem and managed-endpoint coverage described as in preview soon. And it open-sourced Agent Substrate on GKE, an open runtime with the same design that can be tried today on Kubernetes. Google reports for Agent Substrate a 200,000-agent benchmark with 500 activations per second and under 500 milliseconds of suspend/resume latency, vendor figures.
What it means for your business
- Agent infrastructure is now a buying category: sandboxing, identity, access policy, checkpointing and deployment topology are architecture decisions, not details. If your agent vendor cannot explain where its agents run or where credentials live, that is an open question, not a technical detail.
- Cost is won in the pause: an agent waiting on human approval or a slow tool should not bill as if it were working. Ask how idle time is charged before signing.
- Portability versus convenience: Google's managed route (Agent Compute Engine) promises less operations; the open route (Agent Substrate on GKE) promises portability across clouds. The right choice depends on how much control you need and where your systems already live.
Related links
- What to do when your AI agent makes a mistake
- Privacy and data: what happens to your customers' information
- How to implement AI in your business in 30 days (step-by-step plan)
Frequently asked questions
What is Google Agent Compute Engine?
A Google Cloud service announced on October 22, 2026 that gives each AI agent its own hardware-isolated microVM, with no credentials inside the sandbox and pause/resume by snapshot. It already powers Gemini and Antigravity; the customer preview opens in the coming months.
Why does it matter that the sandbox holds no credentials?
Because an agent that runs code and calls tools is an attractive target. If credentials live outside the sandbox and are inserted only in flight, a sandbox compromise does not hand over your keys.
Can I try it today?
Not Agent Compute Engine: its preview arrives in the coming months, with no exact date. Agent Substrate, the open runtime with the same design for GKE, can be tried today on Kubernetes.
What do Google's density and latency figures mean?
They are vendor figures: up to 10x density from shared memory, and 500 activations per second with under 500 ms suspend/resume latency in a 200,000-agent benchmark. Use them as a reference, not a guarantee for your workload.
