Google Open-Sources AX to Orchestrate AI Agents

Google has released AX, an open-source orchestrator and declarative runtime under the Apache 2.0 license, designed to manage autonomous AI agent workloads. Hosted on GitHub as google/ax, the platform runs on top of Agent Substrate and treats agents as stateful actors rather than traditional microservices or batch jobs. This architecture allows AX to address the unique operational demands of AI agents, which often experience bursty reasoning periods followed by long idle phases while waiting for model responses or human intervention.

To prevent compute underutilization during these idle periods, AX checkpoints an agent's execution state and suspends it, achieving sub-second task suspension and resumption with zero cold-start delay. The control plane exposes four declarative Kubernetes-style primitives under the ax.io/v1alpha1 API group: Task, Workspace, Gateway, and Model. The Task primitive defines lifecycles and sandbox resource constraints, while Workspace handles environment setup like mounting Git repositories or configuring Model Context Protocol servers. Gateway manages outbound network security policies, and Model establishes a unified control point for large language model provider parameters and secrets.

Practitioners interact with the system using a Go-based command-line tool called ax. Operators deploy the control plane to Kubernetes using ko and Redis into the ax-system namespace, managing workloads with commands such as ax apply, ax watch, ax ssh, ax suspend, and ax resume. While some developers criticize the heavy operational overhead of managing Kubernetes clusters and custom resource definitions, infrastructure engineers praise the platform for drastically reducing cloud costs associated with idle agents. AX is positioned as a foundational execution runtime utilizing gVisor-isolated sandboxes, making it a robust compute primitive for enterprises managing large-scale agent fleets rather than a high-level application framework.