Google's open-source AX checkpoints idle AI agents so dozens can share a host worker

Agent workloads, in InfoQ's account, compute hard during reasoning, tool execution and local code evaluation, then wait for long stretches on model responses, external APIs or a human ^[3]. A dedicated sandbox kept active through those waits leaves compute unused. Cold-starting a fresh container when the wait ends adds latency to an interactive loop ^[4]. AX runs on a runtime called Agent Substrate and treats each agent as a stateful actor ^[2]. It checkpoints the actor when the wait starts, so the sandbox can give up its share of the host without a cold start when the wait ends ^[5][6].

The waste in that description comes from one pattern: a live sandbox per session, held through every idle phase ^[4]. AX fixes it inside Kubernetes. The control plane is deployed to a cluster with ko and Redis, into an ax-system namespace, and operators reach it through their existing kubectx contexts ^[11]. The Model primitive keeps provider secrets in Kubernetes ^[10].

Both the speed figure and the density figure are stated as design goals ^[6]. The report does not include measured resume latencies or host utilization. For the density to hold on another fleet, its agents have to spend most of their wall-clock time waiting. The state written at each checkpoint also has to be small enough to save and restore inside the resume budget. A coding agent that spends its session compiling and running tests is busy most of that time, and suspension has little to reclaim from it ^[3].

The Gateway primitive is the part I'd copy into any agent runtime. It holds a sandboxed agent to an explicit allowlist of hostnames and ports and injects credentials into outbound requests ^[9]. Egress policy and credentials become a declared resource that sits next to the task. Workspace handles setup. It mounts Git repositories, configures MCP servers, installs skill bundles, or passes a natural-language goal to an initialisation agent that bootstraps toolchains before the task starts ^[8]. A setup step written in English is the one I would want to dry-run first.

Opinion splits over the operational bill. On Hacker News, infrastructure engineers credited AX with tackling the cloud cost of agents idling on model APIs or human input ^[13]. Developers there objected to the project's "ergonomic workflows" pitch, citing the upkeep of Kubernetes clusters, container registries and custom CRDs through tools like ko ^[13]. Posters on Reddit placed AX below frameworks such as LangGraph or CrewAI, as an execution runtime ^[14]. Day to day, operators drive it through a Go CLI. ax apply registers manifests, ax watch streams phase changes, ax ssh opens a sandbox for debugging, and ax suspend and ax resume trigger the checkpoint path by hand ^[12].