Google releases open agentic orchestrator AX for scaling AI workloads
AX provides sandboxed, declarative primitives for running billions of agent tasks, with isolated execution, automatic workspace generation, network policies, model configuration,
Overview According to the source, AX (Open Agentic Orchestrator) is a declarative control plane designed to run agentic workloads at scale. It sandboxes each task, sets up a workspace, fences the network, and can handle billions of tasks per cluster. Users can declare a single task per agent or compose multiple tasks as needed. ## Example Workflow The source provides a sample YAML that defines a workspace pulling the Go repository and a task that ensures the Go toolchain is built from source. After applying the manifest with ax apply -f task.yaml, the workspace and task are created. Commands such as ax watch, ax get tasks, ax ssh, ax suspend, ax resume, and ax delete let users monitor, inspect, interact with, pause, resume, and remove tasks. ## Core Primitives The source outlines four primitives that address the unique needs of agentic workloads: * Isolated execution – run untrusted agent code in a sandbox with CPU and memory limits, cheap to create, suspend, and discard. * WorkspaceEasy workspace setup – list Git repositories, MCP servers, skills, or describe a goal; AX prepares the sandbox before the task starts. * GatewayNetwork policies – define and manage network policies, restricting traffic to an explicit allowlist and injecting credentials. * ModelOne place for config – configure models, parameters, and secrets in one location, allowing quick rotation or version pinning. ## Scaling and Efficiency Built on Agent Substrate, AX runs each task as a lightweight actor, enabling billions of concurrent agent sessions per cluster. Idle agents waiting on model responses, external tool calls, or human input are checkpointed and suspended, then resumed in under a second with zero cold‑start delay. Dozens of tasks share worker resources, turning idle time into spare compute capacity so users pay only when agents are actively thinking or running code. ## Generative Features The platform integrates generative AI directly. Users can describe a desired workspace in plain English, and AX provisions the environment automatically before the task begins. This capability extends to creating Python environments, installing toolchains, and verifying dependencies via generative workspaces. ## Use Cases According to the source, AX supports interactive coding agents, long‑running agent servers, Jupyter notebooks, headless browser testing, and custom tool runtimes. It is especially suited for research, allowing teams to spin up massive numbers of reproducible sandboxes to collect trajectories, run reinforcement‑learning loops, and evaluate agents at scale. ## Design Philosophy For builders and researchers, AX emphasizes ergonomics, rapid iteration, and joyful workflows while keeping the runtime minimal and lightweight. The project originated at Google from agentic runtime systems research; teams recognized that agentic workloads represent a new computing paradigm—stateful, bursty, long‑running actors that compute intensely then wait for model, tool, or human responses. Traditional orchestrators built for stateless microservices or predictable batch jobs become cost‑prohibitive when idle sandboxes remain running and lack native sub‑second suspend/resume capabilities. Drawing on research from Google DeepMind and extensive experience in large‑scale isolation, resumption, and scheduling, AX is being developed as an open declarative control plane purpose‑built for agent execution, abstracting tasks, workspaces, network policies, and models into core primitives while relying on Agent Substrate for the underlying compute runtime.
