AX is a high-throughput, declarative orchestrator designed to run billions of autonomous agent workloads in a cluster. It runs on top of Agent Substrate for sandboxed execution and is built to handle tasks that are neither stateless microservices nor traditional batch jobs, but rather stateful, bursty, and long-running actors.
AX, an Open Declarative Orchestrator for Massive Agentic Workloads
The platform provides four core primitives for managing agent lifecycles: sandboxing, workspace management, network policy enforcement, and model configuration. These primitives allow developers to run untrusted agent code in a sandbox with CPU and memory limits, manage Git repos or MCP servers via workspaces, define network policies to lock down traffic, and configure models and secrets in one place. AX also features generative workspaces that can set up environments based on plain English descriptions.
AX is designed for high density, allowing users to scale to billions of concurrent agent sessions per cluster. It supports sub-second suspend and resume operations, enabling idle agents waiting on model or human responses to be checkpointed and resumed without cold-start delays. The interface is kubectl-shaped, offering commands such as apply, get, watch, describe, and delete to manage tasks, workspaces, gateways, and models.
Sources
- Google's Open Agentic Orchestrator （Hacker News Frontpage, 2026-09-20）
- GitHub