Google has open-sourced AX, an Apache 2.0-licensed platform for running autonomous AI agents whose workloads alternate between intense activity and long waits. Instead of keeping a dedicated container alive, AX checkpoints an idle agent’s state and is designed to resume it in under a second.
The system runs on Agent Substrate, which treats each session as an isolated, stateful actor with CPU and memory limits. Multiple suspended tasks can share host workers, reducing idle compute without forcing agents through a full cold start whenever a model, external API or human responds.
AX exposes four Kubernetes-style resources. Task defines execution and sandbox limits; Workspace assembles repositories, tools and skill bundles; Gateway restricts outbound hosts and injects credentials; and Model centralizes provider settings and secrets. Operators install the control plane on Kubernetes, while developers use a command-line tool to apply, watch, debug, suspend and resume tasks.
The project targets production agents, research trajectories and reinforcement-learning workloads. It adds another infrastructure layer, however, and teams must still secure tool access and decide whether stateful-agent density justifies running AX, Redis and its Kubernetes control plane.