Google has announced Agent Compute Engine, a service that gives every AI agent its own sealed-off virtual computer. Agents are programs that carry out tasks for you. Google says the service already powers Gemini and will open for preview in the coming months.
How it works
Each agent runs in a small virtual machine, which is a software computer living inside a real one. Google says these are secured by its Titanium chip. When an agent sits idle, its machine pauses and frees up room for others. Memory is shared between machines, which lowers the cost of each one.
Google says this lets it fit ten times as many agents on one server. In a test with 200,000 agents, pausing and resuming took under 500 milliseconds, according to Google. The company also open-sourced Agent Substrate, a Kubernetes runtime that sits behind this.
Oversight tools
Google is also adding controls for companies. They include token tracking, which shows how much AI usage an agent racks up, and model routing. It also adds support for outside identity providers such as Ping Identity and SailPoint. The controls will reach Cloud Run, GKE and managed endpoints.
Not yet known: pricing, the exact preview date, and whether the ten times figure holds up outside Google's own tests.
The takeawayIf your business plans to use AI agents, watch for lower running costs and better tracking of what each agent does and spends.