Google has announced Agent Compute Engine, a runtime that places every AI agent in its own hardware-isolated micro virtual machine. The service already supports Gemini internally and is scheduled to enter preview in the coming months.
Each microVM is protected by Google’s Titanium security hardware. When an agent becomes idle, its virtual machine pauses and returns physical resources to the shared pool. Shared memory between virtual machines is intended to lower operating cost without putting agents in the same execution environment.
Google also open-sourced Agent Substrate, a related Kubernetes runtime. The company says it can fit ten times as many agents on one server and that a test involving 200,000 agents suspended and resumed them in under 500 milliseconds. Those are Google’s own performance claims and will need validation under real workloads.
The broader management layer adds token accounting, model routing and third-party identity support from Ping Identity and SailPoint. Google says these controls will extend across Cloud Run, Google Kubernetes Engine and managed endpoints. The announcement provides architecture and preview timing, but no public pricing or exact preview date.