On October 22, Google launched a new service—the Agent Compute Engine (ACE). It is part of Google Cloud and is designed to provide “computers” for AI agents: the goal is to give every agent a secure, durable, and scalable dedicated device.
Unlike traditional virtual machines or containers, ACE is tailored to the lifecycle of each agent. It assigns each agent a hardware-isolated microVM, allowing them to run code, use tools, and browse the browser safely, until they complete multi-step tasks. This environment has an independent client kernel and root privileges, supported by the hardware-isolated Cloud Hypervisor boundary for security. No credentials are stored in the sandbox, reducing the risk of abuse or data leakage for the agents from the outset.
Free from work after hours, woken up in seconds
The most convenient feature of ACE is “persistent and pauseable”. Once the agent has downtime, its computing environment automatically pauses, freeing up resources; when work arrives again, it can quickly restore the file and memory state to their original conditions. According to Google, in testing with 200,000 agents, the pause and resume process takes less than 500 milliseconds. Combined with shared memory and snapshot copying, the number of agents that can be handled by a single server increases significantly.
This engine was initially used for Gemini, the Gemini platform, Antigravity, and Google’s internal model evaluation and reinforcement learning processes. In the future, developers will be able to directly call ACE through Google Cloud, allowing them to control the mirror, identity, network access, and lifecycle of the AI environment, while Google will handle the underlying resource scheduling and scaling.