Google Cloud announces isolated compute service for AI agents

Google Cloud announced Agent Compute Engine on October 22, a managed service designed to run artificial intelligence agents in hardware-isolated microvirtual machines, and said customer preview is planned in the coming months. The service is intended for agents that execute code, interact with browsers and use tools to complete tasks. Google says each agent runs in a microVM with its own guest kernel and root privileges, separated from other workloads by a hardware boundary. The company says its Gemini work agents already run on the service internally. Google says an agent’s microVM can be suspended when idle and restored later from a checkpoint, allowing it to resume its work. The system uses copy-on-write memory sharing, in which agents can share memory until changes require a separate copy. The company says this can reduce resource use when many agents run from a common base. Google has cited benchmarks in which the memory-sharing approach enabled ten times the agent density. That is a company-reported result, not an independently verified customer outcome. In a separate test described by Google, 200,000 agents were suspended and resumed in under 500 milliseconds. The...