Traditional operating systems belong to a tool-oriented architecture, providing file management and runtime environments, and statically responding to mouse or touch commands.
However, as new form factors such as smart glasses and embodied robots emerge, the traditional architecture can no longer meet the demands for high autonomy and real-time perception, thereby driving the evolution of underlying software toward an Agentic Operating System (AIOS).
In this regard, Liang Quan, Vice President of Marketing and Ecosystem at Arm China, pointed out that the future AIOS will be more than just software; AI models will become the core hub of the system architecture.
Figure | Liang Quan, Vice President of Marketing and Ecosystem at Arm China; Source: Arm China
Interaction Paradigm Will Evolve from "Turn-Based" to "Continuous Perception"
The core shift in new hardware lies in breaking the turn-based logic of one question and one answer. Taking embodied intelligence as an example, interactions must be continuous and seamless. Devices need to proactively capture users' movements, eye gaze, and emotions through cameras and microphones, and respond directly, rather than relying on rigid wake words.
Liang Quan believes that this continuous perception will rely on two system shifts: first, Generative UI, where the display interface is no longer fixed but dynamically reorganized and instantly generated by the system based on the current task context and environment; second, multimodal mixed input, where the dataflow engine must simultaneously process video, audio, and text, and feed them instantly to the underlying models for inference.
The Four Core Engines of AIOS
Liang Quan revealed that the current architecture chooses to refactor on top of existing classical OSs such as Linux, sinking AI models and task management into the most fundamental capabilities, driven jointly by four major engines:
- Model Engine: Deeply integrated into the system interaction flow, responsible for instantly parsing subtle emotions in speech and the true intent of text, and performing dynamic routing between the edge and the cloud.
- Agentic Engine: The core lies in loop task orchestration. The system automatically plans and breaks down vague, macro-level single-sentence instructions into several executable sub-tasks, autonomously invoking tools for execution within secure privacy boundaries.
- Dataflow Engine: The data transmission hub of the system. To handle high-frequency video and audio streams, it adopts Zero-Copy technology, enabling data to be transmitted without copying among heterogeneous computing units such as Arm CPUs, GPUs, and the Zhouyi NPU, completing inference and motion control with ultra-low latency.
- Memory Engine: Responsible for long- and short-term memory and context management. Through automatic archiving and retrieval mechanisms, it endows devices with continuous memory, optimizing current decisions based on historical habits.
Feasibility and Engineering Implementation of Edge AI
Liang Quan believes that, benefiting from technological breakthroughs in Mixture of Experts (MoE) and small-scale models (SLM/VLM/VLA), edge models have reached a critical point of high usability when handling local, immediate, and specialized tasks. The generational gap is narrowing, and they possess privacy protection advantages that the cloud cannot match.
To verify system capabilities across various hardware, Arm China and the open-source AIOS alliance will adopt a strategy of starting small and scaling up. The initially developed JishuShell tool appears to be a command-line manager on the surface, but is essentially a prototype of a micro operating system underneath. Optimized for the edge, it features a built-in server-level task orchestration mechanism that can directly allocate AI tasks to run locally or in micro virtual machines. This architecture allows developers to directly test the agentic capabilities of edge models on mainstream hardware such as Mac Studio and NVIDIA DGX SPARK.
As the software infrastructure matures, this technology is expanding from smart recording devices and smart glasses to multi-device interconnected home hubs and embodied robots. Through the refactoring of the underlying operating system, new AI devices are transforming from cloud API tools into independent computing entities capable of continuous perception and autonomous task execution.