Google Cloud unveils Agent Compute Engine for AI agents

Google Cloud has introduced Agent Compute Engine, a specialized execution infrastructure designed to host autonomous AI agents within isolated, virtualized environments. Unlike traditional inference applications that merely generate text, AI agents require the ability to execute code, interact with external tools, control browsers, and manage complex multi-step workflows.

To meet these needs, the service assigns a dedicated micro virtual machine (microVM) to each agentic session. These microVMs utilize a guest kernel and restricted root privileges to ensure isolation via Cloud Hypervisor, preventing code from one agent from accessing the environments of others. The architecture leverages Titanium hardware and Infrastructure Processing Units (IPUs) to strengthen resource separation and operational control.

Key features include the ability to pause an agent's workspace while it waits for model responses, reducing resource consumption without losing operational state. Persistent workspaces are supported through Hyperdisk storage, allowing for the management of memory states and files across different layers. Google plans to release a preview version of the service to customers in the coming months, noting that it is already being used internally for Gemini's agentic tasks.