IT Home On October 23, Google announced yesterday (October 22) that it has launched the Agent Compute Engine (ACE). It is a Google Cloud computing service designed specifically for AI agents, aiming to provide each agent with a secure, persistent, and elastic “dedicated computer.”
According to the official announcement, the ACE service will provide each AI agent with independent hardware-isolated microVMs, allowing it to safely execute code, call tools, control browsers, and complete multi-step tasks.
According to the blog post cited by IT Home, unlike traditional virtual machines or containers, ACE is specially optimized for the lifecycle characteristics of agents. It can start quickly, has an independent client kernel and root permissions, and ensures security through the Cloud Hypervisor boundary with hardware isolation. Google states that no credentials are stored within the sandbox to reduce the risk of agents being abused or data being leaked.
The highlight of the ACE service is “persistent and pauseable”. When the agent is idle, its computing environment can automatically suspend and release resources; when it needs to continue working, it can quickly restore the file and memory status.
Google says, in tests with 200,000 agents, pause and resume took less than 500 milliseconds, and the use of shared memory and snapshot replication can significantly increase the number of agents that a single server can handle.
This service was initially used to support Gemini, the Gemini platform, Antigravity, and Google’s internal model evaluation and reinforcement learning processes. In the future, developers will be able to use ACE through Google Cloud to independently control the image, identity, network access, and lifecycle of the agent environment, while Google will be responsible for the underlying resource scheduling and scaling.