Google’s AX orchestrates billions of AI agent tasks on Kubernetes

Google open-sourced AX, an orchestrator built to run enormous fleets of AI agents, though calling it “Kubernetes for agents” oversimplifies what it actually is. AX doesn’t replace Kubernetes. It runs on top of a standard cluster, submitting jobs as YAML manifests that look and behave like a Deployment or a Job, so anyone who’s already run Kubernetes will recognize the shape of it immediately even though the workloads underneath are entirely different.

The system is built around four primitives. Task spins up an isolated sandbox for each agent with hard CPU and memory limits, backed by Google’s own container-level sandboxing layer called Agent Substrate. Workspace handles setup: cloning the right Git repos, provisioning MCP servers, and loading whatever skills a task needs before the agent even starts working. There’s also a “generative workspace” mode where an agent gets a plain-language goal and scaffolds its own environment from scratch. Gateway locks outbound network access to an explicit allowlist of hosts and injects credentials only where needed. Model centralizes which LLM a task runs against, with secrets stored the same way Kubernetes already stores them.

That combination is really the pitch: isolation, controlled setup, restricted network access, and centralized model configuration, all declared the same way you’d declare any other Kubernetes resource. Developers can shell directly into a running task with an ax ssh command when something needs debugging, which matters once you’re not running one agent but potentially millions running in parallel across a cluster.

Google says AX is designed for billions of agent tasks running in a single cluster, a scale that treats individual agent runs less like long-lived services and more like disposable jobs that spin up, do one thing, and disappear. The project ships under Apache 2.0, still under active development, with the repository itself warning that breaking changes are likely before anything gets called stable. It fits a pattern of platforms racing to formalize agent infrastructure this year, alongside efforts like Unity’s own agent plugin for game development. AX isn’t Google’s only push here either; a separate, related Kubernetes SIG Apps project called Agent Sandbox covers similar ground and shouldn’t be confused with it.