Google has announced Agent Compute Engine, a service that runs each AI agent in its own isolated virtual machine. The service already powers Gemini, and Google says it will open it for preview in the coming months; it has not given a specific date.
Each agent runs in a small virtual machine, or microVM, secured by Google’s Titanium chip, the company says. When an agent is idle, its VM pauses automatically so its machine can be used by another agent. Google also says shared memory between VMs reduces the cost of running each one.
Google’s capacity claims
Google has also open-sourced Agent Substrate, a Kubernetes runtime built on the same approach. The company says it can run ten times as many agents on one server. In a test involving 200,000 agents, suspend and resume took less than 500 milliseconds, according to Google. The source does not describe the test setup or cite independent validation.
For managing agents, Google is adding token tracking, model routing and support for outside identity providers, including Ping Identity and SailPoint. The company says these controls will also extend to Cloud Run, Google Kubernetes Engine (GKE) and managed endpoints. It has not provided further details about availability or pricing.
What’s still unclear
The announcement describes infrastructure for running and overseeing agents, rather than a new agent product for consumers. Google has not said when the preview will begin, who will be eligible, or how much the service will cost.
Why it matters
For developers building AI agents on Google Cloud, separate microVMs and automatic pausing could offer a way to isolate agents while using server capacity more efficiently. The ten-times capacity and speed figures are Google’s claims, and the announcement does not yet specify preview access or pricing, so developers cannot assess the practical cost or availability.
