Step AOS: The First Agent‑Native OS Unveiled at WAIC

Step AOS debut at WAIC

At the World Artificial Intelligence Conference (WAIC) the company unveiled Step AOS, the world’s first operating system built specifically for autonomous agents, together with the personal agent Step Amoo and the hardware platform STEPX Neo.

Why a new OS is needed

Traditional mobile OSes assume a human opens an app and performs step‑by‑step actions. An agent receiving a request such as “plan my Shanghai business trip” must gather calendar data, travel policies, flight preferences, hotel limits, and even arrange airport transfers, then pause for user confirmation at critical points. Keeping the whole workflow inside a single chat window is insufficient.

I need to travel to Shanghai next week; please arrange the trip.

The author’s key question is how Step AOS bridges the gap between “understood” and “fully completed.”

Three‑layer architecture

The system is organized into three layers:

Top layer: NUI (natural user interface) captures user intent.

Middle layer: handles memory, decision‑making, execution, and safety.

Bottom layer: provides hardware compute (CPU, GPU, NPU) and existing Android, Linux, RTOS capabilities.

The middle layer splits memory into a user domain (facts, context, user profile) and an agent domain (domain knowledge, personality, methods). Information follows a “record‑organize‑recall” cycle, with additional consolidation, reflection, and forgetting steps.

Memory management

Long‑term memory must be versioned, source‑tagged, and conflict‑resolved; it should support isolation for multi‑user or multi‑organization scenarios. Users also need visibility into what the system remembers and why it uses that data.

Atomic abilities and function calling

Step AOS decomposes system capabilities into thousands of “atomic abilities” such as reading the calendar, sending SMS, or creating a reminder. In a function‑calling workflow the LLM generates a tool name and parameters; the OS validates permissions and state, executes the concrete API, and returns the result to the model.

Workflow orchestration

Complex tasks are expressed as a graph of Nodes, Edges, and States. Nodes perform concrete actions (e.g., query calendar, filter flights, write itinerary); Edges decide the next step; States persist intermediate results and progress. Persistent state enables the system to resume after a timeout instead of restarting the whole workflow.

Edge‑cloud routing

Simple, high‑frequency actions (set alarm, open app) stay on‑device, while heavy reasoning and multi‑step planning are offloaded to cloud models. The routing policy follows three principles: “run on edge when possible, use cloud when needed, and collaborate edge‑cloud when required.” This impacts latency, cost, and privacy.

Safety and governance

Step AOS adopts four governance pillars: trustworthy, visible, controllable, reversible. A demo shows the agent deleting the last photo in an album; a double‑confirmation dialog appears, and the user can later view the action log and undo the deletion.

Risk‑level controls map tasks to permission levels (allow once, always allow, deny). High‑risk actions require confirmation and idempotent keys to avoid duplicate execution. Audit logs must be de‑identified and retained only as long as necessary.

Conclusion

Step AOS integrates intent capture, memory management, atomic ability orchestration, edge‑cloud routing, and safety controls into a single architecture, turning the mobile device into a platform where LLMs can invoke system resources directly. However, the roadmap is still experimental: long‑term memory accuracy, cross‑app failure recovery, permission fatigue, and audit‑log usability need real‑world validation.