Google Cloud has unveiled Agent Compute Engine, an infrastructure service designed specifically for hosting AI agents. The platform utilizes microVMs, which combine the hardware efficiency of containers with the security isolation of traditional virtual machines to prevent malware spread.
The system operates on stateless appliances that store data in immutable file systems. Security is managed through custom Titan chips for firmware integrity and CPU-level mitigations against side-channel attacks. Access to storage and network resources is strictly limited, with traffic routed through multiple proxies.
To optimize resource usage, the service automatically shuts down idle sandboxes and reactivates them when needed. This approach allows a single physical node to host significantly more agent microVMs than standard virtualization methods, while supporting features like cloning and snapshotting.
Agent Compute Engine manages sensitive credentials outside of the agent sandboxes. A companion service, Agent Gateway, handles authentication and will soon introduce model routing and token management tools to help users track spending and enforce budget limits.
Google plans to release Agent Compute Engine and the updated Agent Gateway in the coming months.
