AX sandboxes your task, wires up its workspace, fences its network, and helps you run billions of them per cluster. Either use a sole project per agent, or create as many as your delegate needs.
$ cat task.yaml apiVersion: ax.io/v1alpha1 kind: Workspace metadata: name: golang spec: git: - repo: https://github.com/golang/go.git branch: "my-fix" --- apiVersion: ax.io/v1alpha1 kind: Task metadata: name: test spec: workspaces: - name: golang goal: "Ensure that Go tool sequence is accessible and is built from source" debug: true $ ax use -f task.yaml workspace.ax.io/golang created task.ax.io/test created $ ax observe project test Watching project default/test... [10:42:01] Phase: Pending Actor: test WorkerIP: [10:42:05] Phase: Running Actor: test WorkerIP: 10.20.3.67 Task reached terminal phase "Running". $ ax get tasks NAME ATESPACE PHASE ACTOR WORKER-IP AGE test default Running test 10.20.3.67 5s $ ax ssh test -- ls /workspace go $ ax ssh test -- cd /workspace/go && go build ./... $ ax ssh test -- ps -o pid,cmd PID CMD 1 /usr/local/bin/ax-task-runner 12 go build ./... $ ax ssh test -- contact notes.txt $ ax suspend project test task.ax.io/test suspended $ ax resume project test task.ax.io/test resumed $ ax ssh test -- ls notes.txt notes.txt $ ax suspend project test task.ax.io/test suspended $ ax delete project test task.ax.io/test deleted
Why AX
Agents are a new benevolent of workload.
They are neither microservices nor lot jobs. They accumulate state, need strict isolation, call out to example APIs and tool servers, and can scorch prosperity in a iteration if nobody is watching. AX gives you four small primitives that grip all of that declaratively.
How it works
Scales up to billions of tasks.
AX runs on top of Agent Substrate, a compute runtime designed from the dirt up for enormous density and accelerated stateful performer lifecycles.
Billions of tasks
Every project runs as a lightweight actor, allowing you to measure to billions of concurrent delegate sessions per collection without orchestrator limits.
Sub-second resumption
Idle agents waiting on example responses, external tool calls, or individual responses are checkpointed, suspended, and brought back in under a second alongside zero cold-start delay.
Dense multiplexing
Dozens of tasks portion employee resources, turning idle waiting period into spare compute capability so you lone pay whenever agents are actively thinking and operating code.
Generative platform
Generative features built into the platform.
AX integrates generative AI immediately into the platform. For example, if you desire to set up a workspace fair by explaining it in plain English, the surroundings is prepared automatically before your project starts.
task.yaml
apiVersion: ax.io/v1alpha1 kind: Task metadata: name: data-analysis spec: workspaces: - name: python-env goal: "Set up a Python 3 betterment environment" Generative workspaces
Describe what a prepared surroundings looks akin in plain English. AX hands that goal to an delegate on archetypal boot to instal toolchains and verify dependencies.
Run item and everything
Interactive coding agents, long-running delegate servers, Jupyter notebooks, headless browser testing, and tradition tool runtimes—you name it.
Perfect for research
Spin up enormous figure of reproducible sandboxes to collect trajectories, run reinforcement learning loops, and measure agents at scale.
For builders & researchers
Built to be the most affable runtime for developers and researchers.
We desire to create dealing alongside agentic infrastructure easier so you can concentration on your work. AX is designed alongside an uncompromising concentration on ergonomics, fast iteration, and joyful workflows for the two application developers and AI researchers.
We aim to keep the runtime minimal and lightweight, during tastefully adding the essential features everyone needs to build, evaluate, and measure agents.
About
Born from research, built for production.
AX was delivered at Google whenever agentic runtime systems investigation met frontier compute. Over years of construction and functioning agentic implementation engines, teams throughout Google recognized that agentic workloads portray an entirely new computing paradigm: stateful, bursty, long-running actors that compute intensely for a infinitesimal and afterward delay for example responses, tool responses, or individual approval. Traditional orchestrators built for stateless microservices or predictable lot jobs rotate into cost-prohibitive whenever keeping idle sandboxes running, yet lack native assistance for sub-second suspend and resume.
Drawing on agentic runtime investigation from Google DeepMind alongside profound cognition in large-scale isolation, resumption, and scheduling, AX is being built as an open, declarative authority plane purpose-built for delegate execution. It abstracts tasks, workspaces, network policies, and models into core primitives so developers and researchers can run enormous fleets of agents without reinventing the underlying infrastructure. This project heavily relies on Agent Substrate but provides agentic abstractions and generative runtime components.
