Google unveils Agent Compute Engine, giving each Gemini agent its own isolated micro VM

A machine of its own for every agent

Google has announced Agent Compute Engine, a runtime environment built specifically for AI agents. In this setup, each Gemini agent runs inside its own micro virtual machine, isolated at the hardware level.

A virtual machine (VM) is a software-defined slice of a computer that behaves like a separate machine. The "micro" label suggests each agent gets a small, lightweight unit rather than a full server.

Why isolation matters

Agents can execute commands on their own, so a failure or compromise in one agent could spill over into other workloads. Separate VMs limit that risk: each agent stays in its own box.

The governance side

Alongside the runtime, Google is expanding governance capabilities that let companies manage their agents across cloud, hybrid and on-premises environments. The focus is not just on running agents but on supervising and constraining them.

Who it is for

The announcement targets teams planning to run large numbers of agents at enterprise scale. As agent counts grow, so does the need for security and oversight, and Google is answering that at the infrastructure layer.