AX – Google’s Open Agentic Orchestrator

You can now deploy agentic workloads that scale to billions of tasks with sub-second suspension and resumption.

Google has open-sourced AX, an orchestrator designed for agentic workloads. AX allows developers to declare agent tasks in a YAML format and run them at scale. The system provides primitives for isolated execution (Task), easy workspace setup (Workspace), network policy management (Gateway), and model configuration (Model). It is built on Agent Substrate, a compute runtime capable of handling billions of tasks concurrently by multiplexing dozens of tasks onto shared worker resources. This design enables sub-second suspension and resumption of agents, making them cost-effective for interactive AI applications. AX can be used for various applications, including interactive coding agents, long-running servers, and research tasks like reinforcement learning loops. The project aims to simplify agent infrastructure for developers and researchers, allowing them to focus on building and evaluating AI agents.

AX introduces a novel approach to orchestrating AI agents, positioning them as a distinct workload category between microservices and batch jobs. Its core innovation lies in managing stateful, bursty, and long-running actors efficiently, addressing the cost-prohibitive nature of traditional orchestrators for such tasks. By leveraging Agent Substrate, AX offers native support for sub-second suspend and resume, which is crucial for interactive AI where agents frequently wait for external responses.

The open-sourcing of AX signifies a move towards standardized infrastructure for the rapidly growing field of agentic AI. Developers and researchers gain access to a purpose-built control plane that abstracts away complex infrastructure concerns, enabling faster iteration and evaluation of AI agents. This could accelerate the development and deployment of sophisticated AI applications that require massive-scale, cost-effective agent execution.