Google announces Agent Compute Engine for AI agents

Google’s October 22 announcement introduces a computing service for AI agents. Each agent receives a hardware-isolated virtual computer with its own operating-system kernel and administrator access. Its attached Hyperdisk storage preserves the workspace and agent context between sessions.

Google’s architecture description separates the controller that calls the model and chooses actions from the sandbox that executes its explicit commands. The sandbox holds no persistent service keys, but does receive a short-lived session identity. For an outgoing request, an external gateway checks user delegation, agent identity, node identity and outbound policies before inserting the backend credential outside the sandbox. The controller remains outside the execution environment; the sandbox sends command output back to it.

The linked Agent Gateway documentation explains what administrators control: a registry lists approved destinations, and access policies determine which agent may reach them. Connections are blocked by default unless an explicit policy grants access. The gateway handles incoming client requests and outgoing agent requests as separate paths; the outgoing path also checks destination registration and sanitizes request content.

That documentation lists Gateway availability on Gemini Enterprise Agent Platform, while support for GKE, Cloud Run and other clouds is still coming through early preview. Feature details and timing remain subject to change.

MarkTechPost also covered Google’s announcement.