Google makes a microVM for each AI agent in its cloud

Google shows Agent Compute Engine, a new Google Cloud system for AI agents. It gives each agent an isolated microVM with hardware that is isolated from the host. The Google post gives this cause: usual virtual machines start too slowly. Containers use the same core as the host. Google uses the system for Gemini agents. The post gives no test results and no release date.

How it works

An agent has 2 parts. The harness calls the model and decides the next step. The sandbox operates the code and the browser. The system knows when an agent works and when it waits. It keeps the work of the agent and does not use more resources than it must.

The numbers

The post gives no speed numbers and no cost numbers. The goal is hundreds of millions of agents. Gautham Thambidorai, a Gemini leader, gives this view: the microVMs give speed, isolated hardware and high density.

What is not known

The text of the post stops before the end of the part on how it is made safe. In the post, keys stay in Titan hardware, and the sandbox holds no credentials. No external group made a test of it.

Insight: Cloud companies compete to own the agent sandbox

Google makes a sandbox where the agent holds no credentials. On 16 October 2026, AWS released Strands Box, an open-source sandbox with a policy language, Dogwood. This reads as a race to control the sandbox where agents operate, because the first sandbox that teams use can set the standard. Teams that select at this time can find it hard to move.