Google / Wikimedia Commons (Public domain)
The new compute service runs AI agents inside hardware-isolated microVMs and is already powering Google's own Gemini agents
Google Cloud has a new home for AI agents, and it comes with locks on every door.
The company unveiled Agent Compute Engine on October 22, 2026, at the AGNTCon + MCPCon events. It is a compute service built specifically for running AI agents at scale.
What Google actually built #
At its core, Agent Compute Engine runs workloads inside hardware-isolated microVMs. Each one gets a dedicated guest kernel and root access.
That isolation matters because agents do things. They run code, call tools and touch data on behalf of users.
Google says the service is tuned for the fast churn and heavy demands typical of agent workloads. In practice, that means agents can start quickly, , and then resume later.
The density trick and the security layer #
The second headline feature is copy-on-write memory sharing.
If many agents start from the same base, they can share the same memory until one of them changes something. Only then does that agent get its own private copy of the altered part. Google says this approach has delivered benchmarks of 10x density, meaning the company claims it can pack far more agents onto the same hardware than it otherwise could.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
On the security side, Google is leaning on its Titanium IPU technology, which the company positions as added protection against adversarial threats.
Google frames the service as part of its Gemini Enterprise Agent Platform. Security and governance for AI agents sit at the center of that platform’s messaging.
Already in use, at least in-house #
Agent Compute Engine is not a lab prototype waiting for its first customer. Google says it already runs the service internally for its own Gemini agents.
Customers will have to wait a little longer. The service is expected to enter a preview phase soon, though Google has not tied that to a firm public date in the announcement.
Why agents need their own infrastructure #
A chatbot answers a question and stops. An agent might browse, write code, run that code, check the result and try again, all without a human clicking anything.
That kind of behavior creates two problems for a cloud provider. The first is scale, because agents can appear and vanish constantly. The second is containment, because an agent with root access in a poorly isolated environment is a risk to everything around it.
Agent Compute Engine is Google’s answer to both at once. Fast boots and -and-resume handle the churn. Hardware isolation and Titanium IPU handle the containment.
What this means #
The 10x density claim is the number to watch. If customers in the preview see similar results, it could meaningfully change the cost of running large fleets of agents.
The other thing to watch is timing. Google has said the preview is coming soon, and how quickly it moves from preview to broad availability will show how ready the service really is outside Google’s own walls.
For now, the most convincing endorsement comes from Google itself. Its Gemini agents are already living in these microVMs, and the company is clearly hoping customers will want to move into the building next. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our