Google has announced Agent Compute Engine, a infrastructure service that provisions individual, hardware-isolated micro virtual machines for autonomous AI agents. Powered by Google’s custom Titanium security chip, the system currently underpins Gemini agent infrastructure and will enter public preview in the coming months, according to the company.
To address the high infrastructure costs of agentic workloads, the platform automatically pauses virtual machines when agents idle, sharing memory across instances to reduce operational overhead. Google reports that the architecture allows servers to host ten times as many agents compared to traditional deployments, achieving sub-500 millisecond suspend and resume times in large-scale tests with 200,000 agents.
Google is also integrating enterprise governance capabilities into the runtime, including token consumption tracking, intelligent model routing, and identity provider integrations with platforms like Ping Identity and SailPoint. These administrative controls will extend across Cloud Run, Google Kubernetes Engine, and managed cloud endpoints.
Why it matters
Enterprise developers get a scalable runtime designed to significantly lower the unit economics of hosting autonomous agents.
Security teams gain hardware-level VM isolation and enterprise identity management for agentic production deployments.
Cloud providers are racing to optimize hardware platforms specifically for high-density, low-latency agent execution.
Source: the-decoder.com



