The narrative of AI is rapidly shifting. If in 2024 people were concerned with the parameter competition of large models, and in 2025 the discussion focused on whose model is more user-friendly, the focus in 2026 will shift to "how to put agents into practical use." This means that users' perception of AI has moved to the action stage, and terminals such as smartphones, PCs, AI glasses, and wearable devices must embrace and undergo the opportunities and pains of transitioning from "AI for Device" to "AI for User." At the recently held Caixin Roundtable, practitioners from fields including chips, AI glasses, AI creation, and consulting agencies jointly outlined the development context of agent terminals: the competition for terminal entry points is intensifying, a distributed computing architecture is imperative, and human-computer interaction is facing the most fundamental transformation in the past three decades.
Agents Redefine Terminal Roles: Who is the Best Entry Point?
The terminal industry is entering a structural adjustment.
First, the traditional growth model driven by hardware parameters and single-point functions is reaching its limits. Coupled with tight memory supply and rising prices, the global smartphone and PC markets are expected to decline by 12.9% and 11.3% year-on-year in 2026, respectively.
Second, the arrival of agents will redefine the value of personal terminals, driving terminals to evolve into L3-level task agents.
Finally, terminal manufacturers are shifting from product providers to ecosystem organizers. Consumers' expectations for terminals are no longer about the improvement of single-product hardware configurations, but rather the overall AI ecosystem experience.
Under this trend, the white paper "From AI Devices to Personal AI: Agent-Driven Terminal Evolution and Industrial Restructuring" introduces the concept of "Personal AI," which refers to a user-centric, agent-embodied form of personal intelligence. Wang Jiping, Vice President of IDC China, stated that Personal AI poses three requirements for terminal evolution: first, providing personalized services that better understand users; second, possessing comprehensive perception capabilities to enable more natural multimodal interactions with users while balancing low power consumption and long standby time; and third, ensuring personal information security.
Beyond capability enhancements, the role of Personal AI terminals differs from traditional terminals. They are not only devices for executing intelligent tasks but also need to become dispatch centers for peripheral computing power and coordination centers across terminals.
Then, among existing or emerging terminal products, which one is the best entry point and core carrier for Personal AI?
Zhu Yuankun, Product and Technology Expert at Qualcomm, leans towards smartphones. As the device users carry with them the longest and use most frequently, smartphones have the most opportunities to understand users and are expected to become personal data centers and mini computing centers in the agent era. "We believe it is unlikely that users will carry a computing power dispatch device while also carrying a smartphone, because the smartphone can largely undertake many functions of this computing device, just as it gradually integrated camera functions in the past."
Zhou Sen, Chief Hardware R&D Engineer at Rokid, "bets" on AI glasses. He believes the best carrier for agents must meet two conditions. First, seamless and accompanying interaction. Seamless interaction refers to high-frequency human interaction methods such as vision, language, and head posture, which correspond to the optical waveguide display, microphone array and speakers, and IMU (Inertial Measurement Unit) sensors in AI glasses, respectively; accompanying interaction means being able to be worn for long periods. In the future, AI glasses could weigh 40 grams or even under 30 grams, meeting the needs for prolonged use. Second, possessing the ability to perceive the world multimodally. The cameras, microphones, and IMUs equipped in AI glasses can continuously generate multimodal perception data streams in real time, allowing agents to truly understand user behavior.
Kazk, the host of "Digital Life Kazk," takes a more radical view: in the future, everyone will have their own dedicated AI. Instead of using hardware as a carrier to extend agents, agents will be the core to coordinate all terminals. "Currently, the core terminal for Personal AI is the PC, because the PC has the most native environment and the most open protocols; as chip and model capabilities advance, the core carrier for Personal AI will shift to smartphones; further down the line, the term 'terminal' needs to be replaced, shifting from hardware as the core carrier to AI as the core carrier. AI is the true entry point."
Device-Cloud Collaboration and Distributed Inference: Building an "Always-On Perception" Computing Ecosystem
Addressing the needs of Personal AI for personalized services, always-on perception, and privacy preservation, the terminal computing architecture is undergoing a new round of reconstruction.
Zhu Yuankun believes that the terminal industry is entering the "Agent AI Era." The foundational modules for building the agent experience include: perception capability, equivalent to the agent's "five senses"; "memory" capability, derived from the analysis and accumulation of perceived information to form a comprehensive understanding of the user; reasoning capability, which involves processing, thinking, and refining existing perceptions and memories; and ultimately, invoking tools for specific task execution based on these foundations.
Currently, the industry's understanding of Agent AI is mostly cloud-based. However, to achieve continuous perception and memory accumulation of users by terminals, it is necessary to enhance device-side computing power and continuous computing capabilities for cross-device collaboration, and to build a distributed architecture coordinating devices, edges, and the cloud. Zhu Yuankun pointed out that device-side and cloud-side do not have a forced division of labor, but rather achieve a natural division of labor and seamless flow based on task characteristics. Device-side AI possesses perception capabilities, holds personal information and memories, and has corresponding planners and locally running models, enabling seamless invocation of cloud capabilities; while the cloud is suitable for complex tasks but is not close enough to the user—"If relying entirely on the cloud, an agent might become an extremely powerful productivity tool, but it would be hard to become a warm agent that truly understands the individual."
Implemented at the chip level, to support the perception, memory, and reasoning stages of Agent AI terminals, as well as the task flow requirements of "cloud-edge-device," "distributed inference" capabilities must be built. First, it must have a Sensing Hub, integrating ultra-long battery life and ultra-low power consumption technical features in architectural design to effectively perceive and identify user scenarios; second is the CPU, as the absolute core of the entire system, requiring strong computing and scheduling capabilities to become the "commander-in-chief" of the system; third is the NPU, undertaking reasoning and thinking tasks on the device side, serving as the "AI computing power core" of the agent; and finally, the modem, which represents connectivity. In the future, based on high-speed connectivity like 6G, the entire system can seamlessly switch and flow tasks across the device, edge, and cloud sides.
When such a computing ecosystem is built, the AI industry is expected to create an AI agent environment that continuously perceives and understands users, achieving "always-on perception." "Human perception does not directly reach the final result; the incremental process of an individual's lifelong memories and experiences is crucial. Only with 'always-on perception' can an agent master the user's chain of thought, truly build a 'you' that knows yourself best, and ultimately serve you. It can even help you reflect, allowing you to become a better version of yourself," Zhu Yuankun said.
From GUI to AIUI: Shifts in Interaction Paradigms and Application Design
If the computing architecture is a "black box," the reconstruction of interaction methods is the change directly perceived by users. Over the past thirty years, from command lines to Graphical User Interfaces (GUI) and then to touch, the underlying logic of interaction design has always been to help humans operate machines more efficiently, but humans have always been the ones adapting to machines. The arrival of agents is reversing this relationship: machines are beginning to proactively understand human intentions and even perform operations on behalf of humans.
Zhu Yuankun believes that the most immediate change in agent terminals is precisely the interaction method. AI input involves intent understanding; even if a user only provides a vague noun, the model can infer their true needs, making human-computer interaction no longer rely on complete and precise commands.
Zhou Sen further pointed out that UI interfaces used to be multiple-choice questions, where users clicked step-by-step within paths preset by developers, whereas AIUI is an open-ended question—users propose needs, and AI generates the interface. Even in terminals worn by users for long periods, agents can generate contextualized interfaces based on their understanding of user habits and preferences. For example, when a long-term investor asks AI about the stock price of a certain stock, AI will directly display the candlestick chart; when a fitness enthusiast wears AI glasses to eat dinner, the glasses will automatically display the food's calories.
This change is far more than simply "replacing typing with voice." When agents can understand intentions, invoke tools, and execute tasks, traditional applications acting as the "intermediary layer" in human-computer interaction are being bypassed—if agents can directly invoke APIs to order food, hail rides, or fill out forms, the value of applications as independent products will face a fundamental re-evaluation. While application developers are refining the interface experience for human users, they must also consider whether they need to design invocation interfaces for agents. The business logic of the entire application ecosystem—from the distribution model of app stores to the monetization paths of applications—could be reshaped.
Kazk gave an example of a 3D printer. In the traditional process, operating a 3D printer requires modeling, part splitting, parameter setting, and adding support structures using Blender or C4D, which has a very high threshold. To allow colleagues to experience 3D printing, Kazk's team encapsulated the 3D printer's operating system into a Skill. Colleagues only need to drag the model into the agent and then tell the agent to invoke the Skill for printing. He frankly stated that future applications will be designed not only for humans but also for agents. On the AI news website he developed, agent users are already ten times the number of human users.
From entry point competition to computing reconstruction and then to interaction paradigm reconstruction, the terminal industry in the agent era is undergoing the most profound transformation since the birth of smartphones. However, on the path from technological breakthroughs to scaled applications, and from innovative experiences to the cultivation of user habits, all parties in the industry still need to continuously refine experiences and enhance ecosystem collaboration, driving agents from "capability demonstration" to "value delivery," and becoming an accessible future for everyone.
Author | Zhang Xinyi Editor | Qiu Jiangyong Art Editor | Malia Supervisor | Wu Lilin