Google open sourced AX, a Kubernetes-style orchestrator for agent workloads, under Apache-2.0, and the repository sits at 7,387 stars

A team running many coding agents in production has been stuck between building the sandbox, resource limits and per-agent network policy in-house or paying a cloud vendor to hide it behind a closed API. Google published AX to github.com/google/ax on 2026-09-22 under Apache-2.0, and the repository stood at 7,387 stars when this item was written. The vendor site is agentexecutor.io and the latest tagged release is v0.3.0 from 2026-09-20.

AX is described on its own site as a "high-throughput, declarative orchestrator to run billions of autonomous agent workloads in a cluster." The docs on agentexecutor.io say it emerged from agentic runtime research at Google DeepMind and builds on execution engines Google runs internally. The homepage still warns that "major breaking changes" are likely before a stable release, so this is a public preview.

Four objects, one runtime

The AX control plane exposes four Kubernetes-style objects, all in the ax.io/v1alpha1 API group. Task defines execution lifecycle and resource limits and runs untrusted code in isolated sandboxes. Workspace pre-wires Git repositories, MCP servers, skill packages and dependencies before an agent starts. Gateway locks the outbound network to an allowlist and injects credentials. Model centralises which LLM the agent talks to, with its parameters and secrets. Operators drive the cluster with an ax CLI written in Go, using verbs like apply, watch, ssh, suspend and resume that read like kubectl.

Underneath sits Agent Substrate, the runtime AX runs on. The vendor site claims sub-second resume from an idle checkpoint with no cold-start delay, along with what it calls "dense actor multiplexing" so one node can hold many agent sessions at once. Those figures come from Google's own site. Neither the repository nor the homepage carries a third-party benchmark yet.

What kinds of workloads

The AX site names four workload types: interactive coding agents, long-running servers, Jupyter notebooks, and browser-based testing. All four fit the same pattern of an agent that spends most of its time waiting for a model response or a human, which is what the suspend and resume calls address. It also mentions a "generative workspaces" mode where an environment is configured from a plain-English description, though the documentation on this is thin.

A platform team weighing this against existing options faces a real trade. Cloud vendors already sell managed agent runtimes with the same primitives, and running your own control plane in a cluster is more operational work. In return, the sandbox, the gateway rules and the model routing sit in your own network, the code is readable, and any behaviour you do not like can be forked. The Apache-2.0 licence is fully permissive.

The pre-stable warning on the homepage means version pinning matters more than usual. A team writing AX manifests today should expect API changes before v1, and should keep the manifest set small until Google signals the surface has settled.

Source

• Google: google/ax on GitHub
• Vendor site: agentexecutor.io
• Coverage: Google Open-Sources AX by Olimpiu Pop, InfoQ