AX – Google’s Open Agentic Orchestrator

AX is a platform that simplifies the management of complex, stateful agent workloads. It provides a set of primitives to handle everything needed for running these agents, from sandbox creation to network policies and model configurations. AX can run untrusted code in a sandboxed environment with defined CPU and memory limits, making it cost-effective and easy to create, suspend, and dispose of as needed.

The platform allows developers to specify the Git repositories, MCP servers, and skills required for an agent, or simply describe the desired goal. It manages network policies, securing traffic with an explicit allowlist of hosts and ports, and injecting credentials into incoming requests. Agent developers can also configure models, model parameters, and secrets in one place, with the ability to rotate keys or pin new model versions with a single command.

AX runs on top of Agent Substrate, a compute runtime designed for high density and fast stateful actor lifecycles. It integrates generative AI directly into the platform, enabling agents to set up workspaces simply by explaining them in plain English. This environment is prepared automatically before the task starts. The platform supports a wide range of applications, from interactive coding agents and long-running agent servers to Jupyter notebooks, headless browser testing, and custom tool runtimes.

AX aims to streamline agentic infrastructure, allowing users to focus on their work. It is designed with ergonomics, rapid iteration, and joyful workflows in mind for both application developers and AI researchers. The runtime is kept minimal and lightweight, while essential features are added tastefully to enable the building, evaluating, and scaling of agents.

AX was developed at Google, emerging from the intersection of agentic runtime systems research and frontier compute. It abstracts tasks, workspaces, network policies, and models into core primitives, providing an open, declarative control plane specifically designed for agent execution.