{"slug": "ax-googles-open-agentic-orchestrator", "title": "AX – Google’s Open Agentic Orchestrator", "summary": "Google released AX, an open-source agentic orchestrator that runs on top of the Agent Substrate compute runtime and is designed to scale to billions of concurrent agent sessions per cluster. AX provides four declarative primitives — Task, Workspace, Gateway, and Model — for sandboxed execution with CPU and memory limits, Git and MCP server workspace setup, network allowlisting with credential injection, and centralized model configuration. The project claims sub-second resumption of suspended agents with zero cold-start delay and dense multiplexing that lets dozens of tasks share worker resources.", "body_md": "# Declare an agentic task. AX runs it at scale.\n\nAX sandboxes your task, wires up its workspace, fences its network, and helps you run billions of them per cluster. Either use a single task per agent, or compose as many as your agent needs.\n\n``` bash\n$ cat task.yaml\napiVersion: ax.io/v1alpha1\nkind: Workspace\nmetadata:\n  name: golang\nspec:\n  git:\n    - repo: https://github.com/golang/go.git\n      branch: \"my-fix\"\n---\napiVersion: ax.io/v1alpha1\nkind: Task\nmetadata:\n  name: test\nspec:\n  workspaces:\n    - name: golang\n      goal: \"Ensure that Go tool chain is available and is built from source\"\n  debug: true\n$ ax apply -f task.yaml\nworkspace.ax.io/golang created\ntask.ax.io/test created\n$ ax watch task test\nWatching task default/test...\n[10:42:01] Phase: Pending    Actor: test               WorkerIP:\n[10:42:05] Phase: Running    Actor: test               WorkerIP: 10.20.3.67\nTask reached terminal phase \"Running\".\n$ ax get tasks\nNAME   ATESPACE   PHASE     ACTOR   WORKER-IP    AGE\ntest   default    Running   test    10.20.3.67   5s\n$ ax ssh test -- ls /workspace\ngo\n$ ax ssh test -- cd /workspace/go && go build ./...\n$ ax ssh test -- ps -o pid,cmd\n  PID CMD\n    1 /usr/local/bin/ax-task-runner\n   12 go build ./...\n$ ax ssh test -- touch notes.txt\n$ ax suspend task test\ntask.ax.io/test suspended\n$ ax resume task test\ntask.ax.io/test resumed\n$ ax ssh test -- ls notes.txt\nnotes.txt\n$ ax suspend task test\ntask.ax.io/test suspended\n$ ax delete task test\ntask.ax.io/test deleted\n```\n\nWhy AX\n\n## Agents are a new kind of workload.\n\nThey are neither microservices nor batch jobs. They accumulate state, need strict isolation, call out to model APIs and tool servers, and can burn money in a loop if nobody is watching. AX gives you four small primitives that handle all of that declaratively.\n\n[Task](https://github.com/google/ax/blob/main/docs/concepts.md#task)\n\n### Isolated execution\n\nRun untrusted agent code in a sandbox with CPU and memory limits. Cheap to create, suspend, and throw away.\n\n[Workspace](https://github.com/google/ax/blob/main/docs/concepts.md#workspace)\n\n### Easy workspace setup\n\nList the Git repos, MCP servers, and skills an agent needs, or just describe the goal. AX sets it all up in every sandbox before the task starts.\n\n[Gateway](https://github.com/google/ax/blob/main/docs/concepts.md#gateway)\n\n### Network policies\n\nDefine and quickly manage network policies. Lock traffic down to an explicit allowlist of hosts and ports, inject credentials to the incoming requests.\n\n[Model](https://github.com/google/ax/blob/main/docs/concepts.md#model)\n\n### One place for config\n\nConfigure models, model parameters, and secrets in one place. Rotate a key or pin a new model version with one apply.\n\nHow it works\n\n## Scales up to billions of tasks.\n\nAX runs on top of [Agent Substrate](https://github.com/agent-substrate/substrate),\n      a compute runtime designed from the ground up for massive density and fast stateful actor lifecycles.\n\n**Billions of tasks**\n\nEvery task runs as a lightweight actor, allowing you to scale to billions of concurrent agent sessions per cluster without orchestrator limits.\n\n**Sub-second resumption**\n\nIdle agents waiting on model responses, external tool calls, or human responses are checkpointed, suspended, and brought back in under a second with zero cold-start delay.\n\n**Dense multiplexing**\n\nDozens of tasks share worker resources, turning idle waiting time into spare compute capacity so you only pay when agents are actively thinking and running code.\n\nGenerative platform\n\n## Generative features built into the platform.\n\nAX integrates generative AI directly into the platform. For example, if you want to set up a workspace just by explaining it in plain English, the environment is prepared automatically before your task starts.\n\n```\napiVersion: ax.io/v1alpha1\nkind: Task\nmetadata:\n  name: data-analysis\nspec:\n  workspaces:\n    - name: python-env\n      goal: \"Set up a Python 3 development environment\"\n```\n\n### Generative workspaces\n\nDescribe what a ready environment looks like in plain English. AX hands that goal to an agent on first boot to install toolchains and verify dependencies.\n\n### Run anything and everything\n\nInteractive coding agents, long-running agent servers, Jupyter notebooks, headless browser testing, and custom tool runtimes—you name it.\n\n### Perfect for research\n\nSpin up massive number of reproducible sandboxes to collect trajectories, run reinforcement learning loops, and evaluate agents at scale.\n\nFor builders & researchers\n\n## Built to be the most friendly runtime for developers and researchers.\n\nWe want to make dealing with agentic infrastructure easier so you can focus on your work. AX is designed with an uncompromising focus on ergonomics, rapid iteration, and joyful workflows for both application developers and AI researchers.\n\nWe aim to keep the runtime minimal and lightweight, while tastefully adding the essential features everyone needs to build, evaluate, and scale agents.\n\nAbout\n\n## Born from research, built for production.\n\nAX was born at Google when agentic runtime systems research met frontier compute. Over years of building and operating agentic execution engines, teams across Google recognized that agentic workloads represent an entirely new computing paradigm: stateful, bursty, long-running actors that compute intensely for a minute and then wait for model responses, tool responses, or human approval. Traditional orchestrators built for stateless microservices or predictable batch jobs become cost-prohibitive when keeping idle sandboxes running, yet lack native support for sub-second suspend and resume.\n\nDrawing on agentic runtime research from Google DeepMind alongside deep experience in\n      large-scale isolation, resumption, and scheduling, AX is being built as an open, declarative control plane\n      purpose-built for agent execution. It abstracts tasks, workspaces, network policies, and models into core\n      primitives so developers and researchers can run massive fleets of agents without reinventing the underlying\n      infrastructure. This project heavily relies on [Agent\n        Substrate](https://github.com/agent-substrate/substrate) but provides agentic abstractions and generative runtime components.", "url": "https://wpnews.pro/news/ax-googles-open-agentic-orchestrator", "canonical_source": "https://agentexecutor.io", "published_at": "2026-09-20 22:32:43+00:00", "updated_at": "2026-09-21 04:52:24.652098+00:00", "lang": "en", "topics": ["ai-agents", "ai-infrastructure", "ai-tools", "agent-protocols", "developer-tools"], "entities": ["Google", "AX", "Agent Substrate", "Task", "Workspace", "Gateway", "Model", "MCP"], "alternates": {"html": "https://wpnews.pro/news/ax-googles-open-agentic-orchestrator", "markdown": "https://wpnews.pro/news/ax-googles-open-agentic-orchestrator.md", "text": "https://wpnews.pro/news/ax-googles-open-agentic-orchestrator.txt", "jsonld": "https://wpnews.pro/news/ax-googles-open-agentic-orchestrator.jsonld"}}