Google Cloud announced Agent Compute Engine on Oct. 22, 2026, outlining a service for running AI agents in hardware-isolated virtual environments. Customer preview access is planned for the coming months, making this an infrastructure roadmap announcement rather than a generally available service launch.
Rapid News Brief — Google says the technology already supports its Gemini agent workloads. Its proposed external offering addresses a different operating pattern from an always-running application: agents repeatedly execute a task, wait for a model or person, then resume work.
Isolation and idle capacity
Google describes small virtual machines, or microVMs, with their own execution environments. It says the architecture can pause idle work and resume it from snapshots, while preserving an agent’s workspace. Those are company descriptions, not independently measured customer savings.
In a separate technical explanation, Google says long-lived service credentials remain outside the sandbox. Outbound requests pass through a gateway that checks identity and policy before supplying the credential needed for the destination service. The distinction is between identifying the execution environment and giving code direct possession of reusable access secrets.
SiliconANGLE reported the announcement and the planned rollout, placing the service alongside updates to Google’s Agent Gateway. Google says new gateway capabilities, including token management and model routing, are planned in preview this quarter. They should not be read as capabilities already available to every customer.
What enterprise teams can evaluate
For businesses planning agent deployments, the relevant questions include what remains persistent when an agent pauses, who authorizes its access and which controls sit outside the code it executes. A sandbox is one component of that operating model, not permission for an agent to take any business action.
Teams should distinguish internal use, preview access and general availability when planning a rollout. They should also test complete workflows rather than assume that faster activation establishes lower total cost. The announcement supplies an architecture and an access timetable expressed in months; it does not establish performance, pricing or suitability for a particular enterprise deployment.