How AX Orchestrates Agent Tasks
AX is an open agent orchestration control plane aimed at running large numbers of stateful AI tasks with bursty compute patterns. Users describe tasks and workspaces declaratively, while AX prepares repositories, tool servers, and environments before execution. Agents then run inside isolated sandboxes.
Its design is organized around four primitives: Task for isolated execution, Workspace for code and dependencies, Gateway for network access and credentials, and Model for centralized model settings and secrets. The demonstrated workflow can create, watch, connect to, suspend, resume, and delete tasks while preserving workspace state. This fits interactive coding, notebooks, browser testing, and custom tool runtimes.
A Runtime Designed for Scale
AX is built on Agent Substrate and treats each task as a lightweight actor. It aims to reduce idle cost by multiplexing worker resources while agents wait for model, tool, or human responses. The project also claims sub-second resumption and a target of scaling to billions of tasks per cluster, supporting large-scale trajectory collection, reinforcement-learning loops, and agent evaluation.
AX’s Main Open Questions
These capabilities are presented mainly as project goals and product claims; the source provides no independent benchmarks, production deployment data, or cost comparisons. Actual resume latency, isolation strength, scheduling scale, model-call costs, and reliability for long-running tasks still need validation through the public code, experiments, and real workloads. The declarative experience is promising, but operational maturity and ecosystem compatibility will determine whether AX can complement or replace existing orchestrators.