Google Unveils AX to Orchestrate AI Agents at Scale

Google has introduced AX, an open-source declarative control plane designed specifically to run and scale AI agent workloads. Developed from agentic runtime systems research at Google DeepMind and other teams within the company, AX addresses the unique computing demands of AI agents. Unlike traditional microservices or batch jobs, agentic tasks are stateful, bursty, and prone to idling while waiting for model APIs, external tools, or human approvals.

AX manages these workloads by treating every task as a lightweight, stateful actor. This architecture allows developers to scale their deployments to billions of concurrent agent sessions per cluster. To optimize resource usage and reduce costs, the system checkpoints and suspends idle agents, resuming them in under a second with zero cold-start delay. Dozens of tasks can share worker resources, ensuring that users only pay for compute power when agents are actively processing code or thinking.

The platform provides four declarative primitives to manage tasks, workspaces, network policies, and models. It runs untrusted agent code inside secure sandboxes with strict CPU and memory limits. Developers can restrict network traffic to explicit allowlists of hosts and ports while injecting credentials into incoming requests. Additionally, AX features built-in generative capabilities. For instance, a developer can describe a desired workspace in plain English, such as asking to set up a Python 3 development environment, and AX will automatically prepare the environment before the task begins.

Built to support interactive coding agents, long-running servers, Jupyter notebooks, and reinforcement learning loops, AX aims to simplify infrastructure management for both researchers and developers. The project relies heavily on Agent Substrate but adds its own agentic abstractions and generative runtime components to deliver a highly ergonomic developer experience.