Google open-sources AX, a Kubernetes-style orchestrator for autonomous agents
Google open-sourced AX under Apache 2.0: agents as stateful long-running actors, sub-second checkpointing, dense per-host multiplexing, on top of Kubernetes.
Google open-sourced AX on September 30 (Apache 2.0, github.com/google/ax): a declarative orchestrator and runtime for autonomous AI agent workloads that treats agents as stateful, long-running actors rather than stateless microservices — filling the scheduling layer that agent infrastructure has been missing.
Facts
- Core abstractions: four declarative primitives — Task / Workspace / Gateway / Model (ax.io/v1alpha1) — with a Go CLI (
ax apply / watch / ssh / suspend / resume) that follows Kubernetes CRD-style usage. - Stateful long-runners: sub-second checkpointing lets idle agents suspend and resume without cold starts — the sharpest difference from container orchestration.
- Dense multiplexing: dozens of sandboxed agents multiplexed per host worker to drive down per-agent cost.
- Deployment: AX is not a container orchestrator itself; it deploys onto existing Kubernetes clusters. Site: agentexecutor.io.
- Sourcing note: the InfoQ report matches the repo docs; the project is young, so production readiness is unproven.
Editorial take
Agent platforms now have open-source parts at three layers: sandboxes (Alibaba’s Agent Sandbox, OpenShell), execution environments (DeepSeek’s DSec), and now Google’s scheduling layer. If your agents are resident and need suspend-without-killing semantics, read the repo before deciding what to build in-house. It is a v1alpha1 API — expect breaking changes; track it, keep it out of production for now.