Google Cloud has introduced Agent Compute Engine, a service designed to isolate AI agents within dedicated microVMs. This architecture powers internal systems like Gemini and Antigravity, separating each agent into a harness for model interaction and a sandbox for executing code or browser tasks.
The platform utilizes Cloud Hypervisor and dm-verity for security, ensuring a read-only root filesystem. By assuming the sandbox is compromised, Google prevents secret theft. The system optimizes resource usage by checkpointing idle microVMs to zero active resources when agents wait for model responses.
Efficiency is achieved through copy-on-write memory sharing and DAX-mapped snapshot images, allowing multiple microVMs to share base OS pages. Storage is divided into disposable scratch overlays, shared read-only volumes, and persistent Hyperdisk workspaces for user data.
The open-source version, Agent Substrate, became available on GKE on September 15, 2026. It supports sandboxes like gVisor and harnesses such as Claude Code and Hermes. While non-production use is open to all GKE customers, production access remains restricted to an allowlist.
Google is also expanding its business governance stack for AI agents. This includes the integration of Agent Gateway, Agent Registry, and Agent Identity across Agent Runtime, Cloud Run, and GKE to provide standardized observability and management tools.
