Google Cloud has unveiled an infrastructure service optimised for running AI agents called Agent Compute Engine, SiliconANGLE reported on Thursday.

Under Agent Compute Engine, AI agents operate in a separate, isolated environment called a sandbox, the report said. The sandbox includes the applications, data and user instructions needed by the agent. It also supports security safeguards to block malicious activity.

Google implemented the Agent Compute Engine sandbox as a microVM. It uses hardware efficiently like a container while offering security comparable to traditional virtualisation, Google said.

Google applied hypervisor technology that isolates virtual machines so malware cannot spread elsewhere even if it gets in.

MicroVMs run on separately configured hardware. The hardware is based on a stateless architecture that does not permanently store files used by customer agents. Stored data is kept on an immutable file system that cannot be modified or deleted. Google verifies firmware integrity with Titan security chips and applied side-channel attack mitigation features to central processing units.

Google plans to roll out Agent Compute Engine with a new version of Agent Gateway in the coming months. The new version adds model routing that sends each request to the model that can handle it most efficiently, and tools to manage agent inference costs. Token management controls token usage across agents and models so it can be used for spending allocation and budget control.