Google’s Agent Compute Engine already powers Gemini work agents; customer preview arrives in coming months with hardware-isolated microVM sandboxes.

Google introduced Agent Compute Engine at AGNTCon + MCPCon as a compute service already supporting Gemini’s work-task agents. External customer availability is scheduled for preview in the coming months, rather than general availability on October 22.

The architecture separates an agent’s trusted harness, which calls models and determines actions, from its sandbox, which executes code, shell commands, tools, and browser interactions. Each agent receives a hardware-isolated micro virtual machine (microVM), with no credentials held inside the sandbox.

Google describes rapid booting, copy-on-write memory sharing across a common operating-system template, and fast checkpoint-and-restore capabilities. Idle microVMs can be paused while an agent waits for a model, person, or external system, then resumed from snapshots.

The service uses Google’s Titanium offload infrastructure, whose infrastructure processing unit (IPU) architecture handles functions including input/output virtualization, storage, and cryptographic trust outside the host central processing unit (CPU). Google also announced expanded Gemini for business governance capabilities for agents wherever they operate, although the supplied excerpts provide no numerical service benchmarks or pricing.