Overview
- Google published AX under an Apache 2.0 license on github.com/google/ax and at agentexecutor.io as a redistributable project.
- AX treats agents as stateful actors and uses gVisor-based sandboxes plus checkpoint and re-hydration to suspend idle sessions and resume them with low latency.
- The control plane exposes four Kubernetes-style primitives—Task, Workspace, Gateway, and Model—under ax.io/v1alpha1 to let teams declare agent lifecycles, environments, egress rules, and model settings.
- Google demonstrated multiplexing 250 agent sessions onto eight pods with sub-second wake times but the company's larger 'billions' concurrency target remains a design goal that others have not reproduced; operating AX requires running Kubernetes and related services via the ax CLI and operator workflows.
- Broad adoption is uncertain because AX faces operational trade-offs versus managed agent APIs, and internal Google alternatives such as Scion and differing team ownership will shape whether AX becomes a durable platform or a niche infrastructure offering.