“The year of the AI agent has arrived.” During the 2026 World Artificial Intelligence Conference (WAIC), Dr. Xu Hao, Global Vice President of Qualcomm and head of R&D for China, made this assessment. As the industry marks 2026 as the “year of the AI agent,” an AI-driven device revolution is moving from concept to store shelves. Honor, Stepfun, Nubia, and other manufacturers coincidentally unveiled their agentic AI phones at this year’s conference, signaling that AI devices with autonomous reasoning and execution capabilities are accelerating from “blueprints on paper” into consumers’ hands.
On July 17, Xu sat down for an exclusive interview with China Business News at WAIC and delivered a speech at the Honor sub-forum. He systematically laid out the fundamental paradigm shift the agentic AI wave is bringing to the device industry, noting that AI is moving from capability demonstrations to more practical application stages. “We are seeing several agentic AI phones being launched in China this year. Compared to previous devices that mainly offered Q&A and assistance functions, AI is now beginning to take on more real tasks, bringing users a more efficient and convenient experience.”
From “Memory” to “Action”: Agentic AI Phones Redefine Human-Computer Interaction
The fundamental difference between agentic AI phones and traditional smartphones lies in the shift in interaction logic—from “humans commanding machines” to “machines proactively serving.” Xu used a series of everyday scenarios to outline the potential of this transformation: users no longer need to open apps one by one and manually input commands; instead, the phone understands intent, breaks down tasks, invokes services, and ultimately delivers results. From “ordering a cup of bubble tea” to “planning a trip,” devices are shifting from passive response to proactive service.
This seemingly simple experience upgrade poses entirely new challenges for device computing architecture. Xu broke down the implementation path from the technical foundation. First, agentic AI devices need to build a “Personal Knowledge Graph” on-device, continuously learning and accumulating user preferences. “For example, what kind of hotels you usually book, what flights you prefer—this personalized information gradually accumulates over time.” Second, devices also need memory management mechanisms similar to human memory, selectively compressing and storing information. “Just like people, we forget things that are unnecessary or unimportant,” creating an efficiently operating personal knowledge base under limited storage and computing power constraints, providing long-term memory and decision support for the AI agent.
At the Honor sub-forum, Xu further pointed out that agentic AI transforms the relationship between people and phones into a three-way relationship among people, devices, and AI agents. Devices need to be able to run anytime, anywhere, and perform sensor processing at extremely low power. For example, when a user is in an indoor meeting, the phone proactively switches to silent mode; when the user steps out of an elevator, the phone automatically handles wireless communication selection. These scenario-based tasks all require always-on, low-power processing modules like Qualcomm’s Sensor Hub to support.
Privacy protection and real-time response demands are driving more AI capabilities to migrate to the device side. “Much of the information involving personal data and requiring very fast real-time processing is better handled on the device side,” Xu stated. Qualcomm’s approach is to build a “cloud-edge-device” collaborative computing architecture—privacy-sensitive and instant-feedback tasks are processed locally, while complex planning loads are handled by the cloud. This collaborative model ensures both a smooth user experience and better balances data security with energy efficiency.
At this year’s WAIC, Honor’s Agentic OS was specifically highlighted by Xu: “This multimodal agentic operating system is one of the best implementations of agentic AI phones we’ve seen.” He emphasized that Qualcomm and Honor have maintained long-term, in-depth cooperation, and in the agentic AI era, they hope to jointly create a new agentic AI-empowered lifestyle.
Three-Layer “Neural Circuit”: Building a Ladder from Perception to Decision-Making for Embodied Intelligence
If agentic AI phones represent AI’s extension into the digital world, then embodied intelligent robots bring AI truly into the physical world. From sorting parts on factory conveyor belts, to service robots fetching and delivering goods in stores, to bipedal robots capable of running marathons, industrial application scenarios are rapidly expanding.
Faced with diverse robot types of varying form factors and task complexity, Qualcomm has proposed a layered computing architecture that divides a robot’s intelligent processing into three levels: real-time perception, motion execution, and logical reasoning.
Xu used a vivid analogy to explain this design philosophy: “We can imagine how humans operate.” The bottom layer is the sensor real-time system, akin to human senses, responsible for low-power, always-on environmental perception—”touching anything, you need to know whether you can grasp it or hold onto it.” The second layer is “cerebellum”-level motion control, responsible for quickly adjusting posture, maintaining balance, and completing coherent movements, with extremely high real-time requirements—”robots need to be able to adjust posture and limb positions at any time to achieve complete balance.” The top layer is the “cerebrum,” responsible for complex logical reasoning, task planning, and command decomposition, namely the VLA (Vision Language Action) model.
This layered design allows Qualcomm to offer differentiated solutions for different robot form factors. Xu explained that different types of robots place different demands on on-device computing. Robot dogs need stable motion and environmental perception; soccer robots need to perform visual recognition, collaborative decision-making, and motion control during high-speed movement; while humanoid robots also need to handle more complex environmental understanding and task planning. To meet these needs, Qualcomm continues to advance its robotics platform capabilities. The newly released Snapdragon IQ10 solution delivers up to 700 TOPS of AI computing power, providing the computational foundation for humanoid robots to operate autonomously in scenarios such as home services and commercial operations.
In Xu’s view, Qualcomm’s long-accumulated experience in high-performance, low-power computing across phones, XR glasses, and smart vehicles is now extending into the robotics field. For different scenarios and tasks, robots need differentiated response mechanisms to balance real-time performance and resource efficiency. “For task scenarios with different response levels, we will have different response priorities,” he said. At the same time, “the automotive industry’s stringent requirements for fault tolerance and millisecond-level response also provide important experience for robotics system design.”
6G and AI Integration: From Communication Pipeline to AI Agent “Neural Network”
As terminal devices become endpoints for AI agents, the role of communication infrastructure becomes increasingly critical. Xu believes that future 6G technology will be the “neural network” that unlocks AI’s full potential, with its design philosophy deeply integrated with AI.
This integration manifests in three dimensions. First, the proliferation of AI applications, especially the high-frequency invocation of device-cloud collaboration by AI agents, will significantly increase data transmission volumes, placing higher demands on network bandwidth and low latency. Second, AI technology itself will increasingly be applied to optimize the design of chips, devices, and network systems, enabling smarter resource scheduling. More importantly, the third point—future 6G will further enhance connectivity and collaboration capabilities, making dynamic distribution of computing power across “device-edge-cloud” possible.
“6G will give us an overall optimization framework from the device side, to the edge, to the cloud,” Xu noted. “When network connectivity capabilities are further enhanced, different computing tasks can be very intelligently distributed between device, edge, and cloud.” This means future terminal devices can make real-time decisions on where to process data based on task nature, privacy level, and computing power requirements. This flexibility will provide important support for the next phase of large-scale AI application development.
Agentic AI Device Ecosystem Accelerates
Looking back at this year’s WAIC theme of “agentic AI deployment,” an important industry trend is visible: more and more terminal devices are gaining the ability to understand, decide, and execute. Beyond agentic AI phones, smart vehicles and embodied intelligent robots are also seen by Xu as key tracks.
“China’s new energy vehicle industry has many leading innovative applications, with many generative AI or agentic AI algorithms already applied in cars,” he said. “In the future, cars will not just be transportation tools; they will gradually evolve into mobile intelligent living spaces.” In the agentic AI phone space, Honor’s Robot Phone and Agentic OS, Stepfun’s STEPX Neo, and Nubia’s officially announced NaviX Ultra—three products making a collective debut—signal that the pace of agentic AI devices moving from “blueprints on paper” to “physical products in hand” is faster than most anticipated.
Xu concluded the interview by summarizing that in the coming years, fields such as agentic AI phones, smart vehicles, and embodied intelligent robots are all expected to see sustained innovation and rapid development. Qualcomm is building an agentic AI-centric device ecosystem through heterogeneous computing architectures, on-device perception capability optimization, and broad device product coverage. From smart earbuds to data centers, Qualcomm has a range of platforms covering various device form factors, capable of supporting all product types in evolving to the agentic AI support stage.
From agentic AI phones to smart vehicles, and on to embodied intelligent robots, AI is accelerating its integration into various scenarios of people’s work and life. And as device computing, communication connectivity, and artificial intelligence technologies continue to advance, an intelligent world composed of countless AI endpoints, seamlessly blending the physical and digital, is moving from WAIC’s exhibition booths into real life.