Google AX: Detailed Introduction

[ikaijua]

Google AX: Detailed Introduction

AX is Google's open-source agentic orchestration runtime — positioned as "the Kubernetes for agents". Created on 2026-03-30, it has grown to 12.2k+ stars and is under active development (breaking changes are expected before a stable release).

What it does

• Define agent workloads with declarative YAML manifests instead of imperative code.
• Runs each agent task inside a sandbox with resource isolation and CPU/memory limits.
• Lifecycle controls: ax suspend/resume for pausing long-running tasks, ax ssh to drop into a sandbox for debugging.

Core abstractions

• Workspace: a pre-packaged environment (Git repos, MCP servers, skill packs) that tasks run against.
• Task: an agent task executed inside a sandbox, with configurable CPU/memory limits.
• Model: the LLM configuration used by the platform itself.

Architecture

AX runs on top of Google's Agent Substrate and is designed around a single cluster that targets billions of autonomous agent workloads.

Quick start

Install the ax CLI and follow the getting-started guide in the official repository README: https://github.com/google/ax