Google has unveiled its new Agent Compute Engine (ACE), a dedicated Google Cloud service designed to provide AI agents with their own secure, persistent, and scalable computing environments. This innovation aims to equip each AI agent with a dedicated “desk” – essentially, its own isolated virtual machine – allowing for enhanced performance and security.
Dedicated Hardware for AI Agents
ACE offers each AI agent a distinct hardware-isolated microVM. This isolation ensures that agents can securely execute code, utilize tools, control browsers, and perform complex, multi-step tasks without interference. Unlike traditional virtual machines or containers, ACE is specifically optimized for the unique lifecycle of AI agents. It boasts rapid startup times, independent kernel access, and root privileges, all secured by hardware-isolated boundaries managed by Cloud Hypervisor.
Security is a paramount concern, and ACE addresses this by ensuring that credentials are not stored within the sandbox. This significantly reduces the risk of agent misuse or data breaches.
“Persistent and Pausable” Computing
A key feature of the ACE service is its “persistent and pausable” nature. When an AI agent is idle, its computing environment can automatically suspend, freeing up resources. When the agent needs to resume its work, its state, including files and memory, can be quickly restored. This dynamic resource management ensures efficiency and cost-effectiveness.
Google reports impressive performance metrics from tests involving 200,000 agents. The suspend and resume operations took less than 500 milliseconds. Furthermore, the service leverages shared memory and snapshot copying to dramatically increase the number of AI agents that can be supported on a single server.
Internal Use and Future Developer Access
Initially, the ACE service has been instrumental in supporting Google’s internal operations. It has powered Gemini, the Gemini platform, Antigravity, and various internal model evaluation and reinforcement learning processes. This internal validation demonstrates the robustness and capabilities of the ACE infrastructure.
Looking ahead, Google plans to make ACE available to developers through Google Cloud. This will empower developers to have granular control over their AI agent environments, including managing the agent’s image, identity, network access, and lifecycle. Google will handle the underlying resource scheduling and scaling, allowing developers to focus on building and deploying their AI agents.
The Future of AI Agent Computing
The introduction of Agent Compute Engine marks a significant step forward in how AI agents are developed and deployed. By providing dedicated, secure, and efficiently managed computing resources, Google is laying the groundwork for more sophisticated and powerful AI applications. The rapid wake-up times and resource elasticity offered by ACE suggest a future where AI agents can be deployed more dynamically and cost-effectively, paving the way for broader adoption and innovation across various industries.
