AX: Google’s Open Agentic Orchestrator

AX is a specialized orchestration layer designed to sandbox tasks, configure workspaces, secure network perimeters, and enable the deployment of billions of agents per cluster.

Whether you require a single-task agent or a complex composition of multiple tasks, AX provides the necessary infrastructure.

🚀 Quick Start Guide

To begin, define your environment and goals in a configuration file.

Example: 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 chain is available and is built from source"
  debug: true

Deployment and Monitoring:

$ ax apply -f task.yaml
workspace.ax.io/golang created
task.ax.io/test created

$ ax watch task test
Watching task 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".

Managing the Agent: You can inspect the state, execute commands via SSH, and control the lifecycle of the task.

Lifecycle Control:

$ ax ssh test -- touch notes.txt
$ ax suspend task test
task.ax.io/test suspended

$ ax resume task test
task.ax.io/test resumed

$ ax ssh test -- ls notes.txt
notes.txt

$ ax suspend task test
task.ax.io/test suspended

$ ax delete task test
task.ax.io/test deleted

---

🧠 Why AX? A New Paradigm for Workloads

Agents are not typical applications. They maintain persistent state, require rigorous isolation, interact with external model APIs/tool servers, and possess the potential to incur massive costs if left in an unsupervised loop.

AX addresses these challenges through four declarative primitives:

1. Task: Provides isolated execution for untrusted code using strict CPU and memory constraints.
2. Workspace: Simplifies environment setup by defining Git repositories, MCP servers, and required skills—or simply by stating a goal.
3. Gateway: Manages network policies. It restricts traffic to a specific allowlist of ports and hosts and handles credential injection for requests.
4. Model: A centralized hub for configuration, managing model parameters and secrets.

---

⚙️ The Engine: Agent Substrate

AX is powered by Agent Substrate, a runtime engineered for extreme density and rapid stateful actor transitions.

Key Technical Advantages:

• Massive Scalability: By treating every task as a lightweight actor, AX can support $\text{billions}$ of concurrent sessions per cluster without hitting orchestrator bottlenecks.
• Instant Resumption: Agents waiting for human input or model responses are checkpointed. They can be resumed in $< 1$ second, eliminating cold-start delays.
• Resource Efficiency: Through dense multiplexing, multiple tasks share worker resources. This ensures you only pay for active computation: $\text{Total Cost} \approx \sum (\text{Active Thinking Time} + \text{Code Execution Time})$
• Generative Infrastructure: AX can build environments from natural language.

Example: Generative Workspace

apiVersion: ax.io/v1alpha1
kind: Task
metadata:
  name: data-analysis
spec:
  workspaces:
    - name: python-env
  goal: "Set up a Python 3 development environment"

In this scenario, AX assigns an agent during the first boot to install the necessary toolchains and verify dependencies based on the English description.

---

🛠️ Versatility and Vision

AX is designed to handle a vast array of agentic needs:

• Interactive coding assistants
• Persistent agent servers
• Headless browser automation
• Jupyter notebook environments
• Custom tool runtimes
• Large-scale RL loops and trajectory collection for research

For the Builders & Researchers

AX prioritizes ergonomics and joyful workflows. The goal is to remove the friction of infrastructure management so researchers can focus on agent behavior.

The Evolution of Orchestration: Stateless Microservices $\rightarrow$ Predictable Batch Jobs $\rightarrow$ Stateful, Bursty Agentic Actors

Born from the intersection of Google DeepMind's research and frontier compute, AX recognizes that agents compute intensely for short bursts and then idle. Traditional systems are either too expensive (keeping sandboxes idle) or too slow (lacking fast resume). AX solves this by providing an open, declarative control plane that abstracts the complexity of isolation and scheduling.