On October 22, Google launched the Agent Computing Engine (ACE), providing AI agents with a secure, persistent, and horizontally scalable dedicated computing environment powered by Google Cloud. ACE utilizes hardware-isolated micro virtual machines, enabling wake-up and state restoration in seconds. In stress tests involving 200,000 agents, pause and resume times remained under 500 milliseconds. ACE is designed for use with the Gemini platform and internal AI workflows, and in the future, developers will be able to access it directly through Google Cloud.Author and source: AIBase
On October 22, Google launched a new service—Agent Compute Engine (ACE), under Google Cloud—designed to build dedicated "computers" for AI agents, aiming to give each agent a secure, persistent, and horizontally scalable专属 device.
Unlike traditional virtual machines or containers, ACE is specifically designed around the lifecycle of agents. It assigns each agent a hardware-isolated microVM, enabling secure execution of code, tool invocation, and browser operations until multi-step tasks are completed. This environment features an independent guest kernel and root privileges, with security reinforced by the hardware-isolated Cloud Hypervisor boundary. No credentials are stored within the sandbox, fundamentally minimizing the risk of agent misuse or data leakage.
Idle after work, wake up in seconds
What makes ACE most clever is its "persistent and pauseable" capability. Once an agent becomes idle, its computing environment is automatically suspended, freeing up resources; when work arrives, it can quickly restore files and memory states to their original condition. Google revealed that in stress tests involving 200,000 agents, the entire pause and resume process took less than 500 milliseconds. Combined with shared memory and snapshot replication, this significantly increases the number of agents that can be hosted on a single server.
This engine originally served Gemini, the Gemini platform, Antigravity, and Google's internal model evaluation and reinforcement learning workflows. Going forward, developers can directly access ACE via Google Cloud to independently manage agent environment images, identities, network access, and lifecycles, while Google handles the underlying resource orchestration and scaling.