Google AX GitHub: Google’s New AI Agent Orchestrator

Looking for the Google AX GitHub repo and wondering what it actually is? The quick version is just below — I’ve been through the README and quick start at github.com/google/ax (fetched 25 September 2026), and this is Google quietly shipping a Kubernetes-for-agents.

Google AX GitHub: what the repo actually is

Per the official README at github.com/google/ax (as of 25 September 2026), AX is “a high-throughput, declarative orchestrator to run billions of autonomous agent workloads in a cluster”. It runs on top of Agent Substrate for sandboxed execution, and Google says it plainly: “If you have used Kubernetes, ax will feel similar.”

The reasoning in the README is the best short argument I’ve seen for why this category exists. Agents are a genuinely new kind of workload: they’re neither stateless microservices nor run-to-completion batch jobs. They accumulate state, need strict isolation, call out to model APIs and tool servers, and — Google’s words — “can burn money in a loop if nobody is watching”. AX answers that with three declarative primitives:

Primitive	What it does
Task	Runs untrusted agent code in an isolated sandbox with CPU and memory limits
Workspace	Pre-wires Git repos, MCP servers and skill packages so every agent starts warm
Model	Configures which LLM the platform uses, with credentials from a Kubernetes secret

Everything is expressed as ax.io/v1alpha1 YAML manifests and applied with a single command. On top of that you get lifecycle commands that solve real agent-ops pain: ax suspend and ax resume checkpoint an idle agent and pick it up exactly where it left off, and ax ssh lets you shell into a running agent’s sandbox to see what it’s actually doing — looking over the agent’s shoulder, literally.

Inside the Google AX GitHub quick start

The install path from the README, condensed:

• CLI: go install github.com/google/ax/cmd/ax@latest — puts the ax binary in your Go bin directory.
• Control plane: you need a Kubernetes cluster, the ko build tool, a container registry your cluster can pull from, and a reachable Agent Substrate Control API. Then make deploy AX_IMAGE_REPO=<your-registry> deploys Redis plus the control plane images into the ax-system namespace.
• First task: ax apply -f examples/task.yaml, then ax get tasks, ax watch task to stream phase changes live, and ax ssh to poke around inside the sandbox.

A minimal manifest pairs a Workspace (say, a Git repo on a branch) with a Task that states a goal in plain English and can enable debug: true so you can shell in. It’s a genuinely small API surface for what it does.

One honest caveat before you get excited: the README opens with a warning that Google is “still actively refining our core concepts, protocols, and specifications” and will “likely introduce major breaking changes prior to a stable release”. This is v1alpha1 territory. Evaluate it now, but don’t move your production agent fleet onto it this week.

Why the Google AX GitHub launch matters

The momentum is real: the agents-radar open-source tracker logged +2,305 stars for google/ax on 23 September 2026 alone, one of the sharpest single-day surges of any agent-infrastructure repo this month. That’s not retail hype — that’s platform engineers bookmarking the thing they expect to be evaluating next quarter.

The bigger signal is Google putting institutional weight behind agent orchestration as an infrastructure layer. We’ve watched this movie before: frameworks proliferate, then the hyperscaler ships the boring, declarative, cluster-native version and the ecosystem consolidates around it. AX sits at a different altitude from application frameworks like the Microsoft Agent Framework or routing layers like 9Router — it’s the layer that runs the agents those tools define. And because Workspaces pre-wire MCP servers and skill packages, it plugs straight into the same ecosystem your Claude Code agents already use.

For those of us running AI SEO operations, the medium-term takeaway is capacity: sandboxed, suspendable, resource-capped agents at cluster scale is exactly the substrate you need to run hundreds of site audits, content refreshes and internal-linking passes concurrently without one runaway loop torching your API budget. That “burn money in a loop if nobody is watching” line was written by someone who has felt that pain — and inside the AI Profit Boardroom, cost-capped agent design is one of the things we drill constantly.

The bottom line on Google AX GitHub

The Google AX GitHub repo is Google’s opening bid to become the Kubernetes of AI agents: three clean primitives (Task, Workspace, Model), YAML-declarative, sandboxed by default, with suspend/resume and ssh-into-the-agent built in. It’s free, it’s open source, it’s trending violently — and it’s explicitly not stable yet, so treat it as a strong evaluation candidate rather than a production bet.

Want to figure out what agent infrastructure your SEO operation actually needs before you commit engineering time? Book a free SEO strategy session and we’ll scope it against your real workload.

FAQ: google ax github

What is Google AX on GitHub?

AX (github.com/google/ax) is Google’s open-source, high-throughput declarative orchestrator for running autonomous agent workloads in a cluster. You declare agentic tasks in YAML, and AX sandboxes them, wires up their workspaces and runs them at scale — Kubernetes-style, but for agents.

How do I install Google AX?

Install the CLI with: go install github.com/google/ax/cmd/ax@latest. For the control plane you need a Kubernetes cluster, the ko build tool, a container registry your cluster can pull from, and a reachable Agent Substrate Control API, then run make deploy with your registry set.

Is Google AX free?

The code is open source and free to use from the GitHub repo. Your real costs are the infrastructure it orchestrates — the Kubernetes cluster and the model API calls your agents make.

Does Google AX require Kubernetes?

Yes, for the control plane. AX deploys into a Kubernetes cluster (everything lands in the ax-system namespace, with Redis) and runs on top of Agent Substrate for sandboxed execution. The manifests will feel instantly familiar if you know Kubernetes.

What are Task, Workspace and Model in Google AX?

The three core primitives. A Task runs untrusted agent code in an isolated sandbox with CPU and memory limits; a Workspace pre-wires Git repos, MCP servers and skill packages so agents start warm; a Model configures which LLM the platform uses, with credentials from a Kubernetes secret.

Is Google AX production-ready?

Not yet, by Google’s own admission. The README carries a warning that core concepts, protocols and specifications are still being refined and that major breaking changes are likely before a stable release. Treat it as something to evaluate, not something to bet production on this week.

Related reading

• Microsoft Agent Framework 1.19.0
• 9Router GitHub Guide
• Claude Code AGENTS.md Support

Last updated September 2026. This is the living guide to google ax github — it gets updated as the tools change.