85. Five presidents from Ant, Qwen, Geely, OPPO and vivo discuss Agent’s first commercial year, funnels and amplifiers

Summary

• The roundtable centered on 2026, widely seen as Agent’s first commercial year, with the core tension captured by Ant CEO 韩歆毅: “There isn’t enough traffic anymore, and traffic is too expensive” （“流量不够了，流量太贵了”）。 Touch, production and fulfillment efficiency have improved sharply over the past 20-30 years; the missing piece is information-conversion efficiency, because “intelligence” is still absent. Host 卫诗婕’s framing was that the traffic era was about funnels, while in the Agent era AI is the “amplifier,” creating new matching efficiency, commercial roles and value distribution—the latter three being 韩歆毅’s views as relayed by the host.
• OPPO’s 唐凯 delivered the sharpest call of the roundtable: “Our model capabilities have already surpassed the organizational capacity of the entire industry chain.” Building an Agent with a good user experience is not difficult, but data and services remain “full of islands”; the industry needs consensus and coordination, and “sometimes even has to fight.” 韩歆毅 went further: industry organization may, candidly, never catch up with model capability, but the pace of technological change “could be faster than any of us imagine”—the shift from OpenClaw to Harness took only months, while DeepSeek-V3 to this year spans no more than 20 months.
• Devices are shifting from being “app entry points” to “intent entry points.” 唐凯’s three-part thesis: devices become intent entry points, always-on sensors with full-domain memory, and multi-device executors. 吴嘉 of Qwen’s end-state view is that users will ultimately remember “an assistant that can provide integrated services across devices,” not one tied to any particular terminal; it also should not be a closed super app, but an open, interconnected agent network.
• 韩歆毅 believes the bottleneck in Agent adoption “is not capability or technology itself, but the mechanism for sharing the benefits across the entire system behind it.” A new mechanism is unlikely to be designed by one company from a god’s-eye view; it is more likely to be hammered out and built jointly by ecosystem participants. 吴嘉 added the infrastructure point: it is “extremely important” that all payments, calls, transactions and fulfillment in the ecosystem can be recorded and traced; otherwise, “we won’t know how to run this ecosystem.”
• Trust already has proof points: OPPO and Alipay repeatedly debated whether 小布’s invocation of Alipay should jump users into another app—“after all, that’s one DAU”—before ultimately letting technology, data and user experience decide. In the first half, Alipay’s DAU on OPPO-branded phones “should have increased by more than 2M,” 唐凯 concluded: “Putting aside what we already own and have gained, and building the future together, is the first step toward trust.” vivo’s 周围 provided corroboration: 蓝心小V answers health-related questions “unconditionally,” with no commercial partnership discussed at any point.
• vivo’s 周围 gave the roundtable’s most measured estimate for next year’s user experience: “walking on thin ice.” Harness’s productivity explosion on PC rests on complete system adaptation to more than 30,000 Skills; that work has only just begun on phones, “not one of those 30,000 can be missing,” and will take at least a year. 任向飞 of Geely believes 10-million-scale atomic Agents could usher in an “Agent internet era,” forcing automakers to develop dedicated AI OSs.
• 韩歆毅, prefacing his forecast with “let me be bold,” said some merchants may begin to feel Agent’s value within 3 months, more may join by around 6 months—roughly after next year’s Spring Festival—and the inflection point, as well as the true breakout, could both arrive in the second half of next year. The host also noted the steady reports that Apple Intelligence will connect to Alibaba’s Qwen foundation model, calling it an important handshake between terminals and models.

Deep dive

1. Traffic at a ceiling: intelligence closes the final gap in information conversion efficiency

• Host 卫诗婕’s starting premise: mobile-internet traffic has peaked. Once an Agent can identify intent like a human assistant and complete tasks end to end, “in many cases people will no longer need to open an app.” The traffic era was about funnels; the Agent era makes AI an amplifier, magnifying both personalized demand and commercial opportunity.
• As relayed by the host, 韩歆毅 sees 3 changes once machines truly understand human language: new matching efficiency, new commercial divisions—including new roles, professions and industry chains—and new value distribution. His commercial diagnosis was blunt: “There isn’t enough traffic anymore, and traffic is too expensive.” Efficiency across the commercial chain has improved over the past 20-30 years; information conversion is the lone remaining gap, and “intelligence can fill this final link.”

2. The essence of service has changed: from silos to organization around the foundation model

• 吴嘉’s framework: in the mobile-internet era, websites, apps and mini-programs delivered services in different ways. In the AI era, services will be “largely unified around the foundation model as the core for delivery and organization,” creating 3 sources of value: a much higher ceiling for service capability, a sharply lower usage barrier through natural language, proactive service and autonomous execution, and a “transformational” improvement in execution efficiency.
• 任向飞 treated his own presence at the event as evidence: “My sitting here and participating in this event should itself represent a fundamental change.” Geely was the first automaker in the sector to propose turning the car from a mobility tool into an “intelligent lifeform,” and the first to lay out an all-domain AI strategy. He recalled watching Knight Rider as a child: “Back then, people imagined a car with a form of artificial intelligence… I feel the time may have finally arrived.” Agent technology, he said, could be a crucial path in that transition.

3. The real anxiety: industry organization lags model capability, and hardware has a September deadline

• 唐凯’s headline observation: “Model capability has now surpassed the organizational capacity of the entire industry chain.” Agents need data to understand users and service tools to create value, but both remain “full of islands.” The industry still needs consensus and coordination, and “sometimes even has to fight.”
• 周围’s 7-8-year timeline: Transformer arrived in 2018, yet voice assistants “still seemed pretty dumb”; 2023 brought foundation models, 2024 saw the entire industry discussing prompt engineering, and Google I/O in 2025 was all about context engineering. Then came Harness on March 11, 2026—“if I remember correctly”—and “all in, it has only been five months and a few days. I think those five months and a few days have produced the biggest change in technology and experience across this 7-8-year period.”
• The hardware deadline is hard: “The phone industry generally hands in its homework in September.” From March to now, that leaves only about 6 months. 周围’s response is not that the pressure is keeping him awake, but that “there doesn’t seem to be much time left to sleep.” The constant is the original mission: use AI to build continuous personalization and give users a personal assistant.

4. The terminal’s 3 shifts: intent entry point, sensor and multi-device executor

• 唐凯’s three-part thesis: the terminal is currently an app entry point, but will become an intent entry point. Users want the final service outcome; API, APK and Skills are merely routes. As intent capability improves, Agents will handle more orchestration and even decision-making. The terminal will then become a more powerful sensor with “always-on perception and full-domain memory,” before ultimately serving as a multi-device executor, with glasses, audio and health devices forming a distributed intelligent-agent system that evolves toward “giant intelligence.”
• 周围 added the hardware foundation: during the transition, the industry needs to integrate compute chips and move foundation models on-device. “A giant model such as DeepSeek-V4 Pro needs to be MoE-ified and made capable of running on phones; that may arrive within another 1-2 years.” The other priority is co-developing the A2A protocol and industry standards with Alipay.
• 任向飞 identified the ecosystem deadlock: when an automaker’s primary Agent tries to call other Agents, “the Agent itself is also a relatively closed island.” Every player has its own ecosystem, creating a natural obstacle to end-to-end delivery. AI needs coordination across society and the entire ecosystem, “but the ecosystems are themselves relatively fragmented.” In the near term, that contradiction will be a major challenge for terminal Agents, especially deeper Agent adoption.

5. 吴嘉: the assistant should not be tied to a terminal, and the platform should not be a black box

• Qwen’s strategic choice is to build its own solution for AI glasses and keep it open because the form factor occupies a unique position—“right where a person’s eyes are.” In mature terminal categories, it will work with manufacturers; “the core is still the model and the Agent.” Looking ahead, users will remember an assistant that provides integrated services across devices rather than one tied to a specific terminal, alongside vertical assistants such as 阿福.
• The fundamental difference between the AI era and mobile internet is that the latter sought longer user time and more information, until incremental gains eventually disappeared. In the AI era, “the longer the time, the more dimensions and the more terminals, the more incremental value this service will continue to generate”; the qualitative change keeps compounding. The service must also avoid being reduced to a pipe: “This is not merely a technology issue; it is also an extremely complex ecosystem and commercial-design issue.”
• The open platform makes 3 commitments. First, subject to privacy, it will share context fully with partners: “Extreme personalization is the defining service characteristic of the AI era.” Without it, efficiency remains an average, which is still the mobile-internet model. Second, it will tell users which Agents were crossed, which Skills were called, and how authentication and review work, avoiding a black box while allowing users to modify and reorder the process with instructions. Third, users can configure the Agents they trust.
• The infrastructure work began last November, when Qwen App moved from information services toward task execution. Qwen worked with Alipay and Alibaba Group on AIF, accounts, credit and related systems. It is “extremely important” that every ecosystem action—including payment calls, transactions and fulfillment—can be recorded and traced; otherwise, “we won’t know how to run this ecosystem.”

6. 韩歆毅: the challenge is not technology, but the mechanism for sharing returns

• The core argument: “The real challenge may be that, in a more open ecosystem, we need to build a new mechanism through which everyone can share the returns.” Network effects will become stronger in the AI era, making it difficult to imagine a single terminal or super app encompassing everything in the foreseeable future.
• Alipay is moving down the stack to build the ecosystem’s most-needed capabilities: trust, payments and parts of security. The new allocation mechanism “may not be exactly the same” as the traffic-and-time model. Will one party design it from a god’s-eye view, or will the participants develop it through friction and iteration? “I think it should probably be the latter.” That path is challenging and highly uncertain, but without a way to share returns, it will be difficult to bring everyone together around the ecosystem.

7. Trust in practice: the app-switch debate and 2M additional DAU

• 唐凯 described how the A2A protocol was born. Early in the partnership, the two sides repeatedly debated whether 小布’s invocation of Alipay services should jump users into another app. “At first, we insisted on jumping out, because after all, that’s one DAU.” The eventual consensus was to let technology, data and user experience decide whether to jump, and the broader market grew instead: in the first half, Alipay’s DAU on OPPO-branded phones “should have increased by more than 2M.” His conclusion: “Putting aside what we already own and have gained, and building the future together, is the first step toward trust.” A2A should resolve many underlying security issues while enabling a smoother experience, including no second login.
• 周围 offered a parallel example: when users ask 蓝心小V health-related questions, “we unconditionally provide an answer.” No commercial partnership was discussed throughout the process—“because over all these years, he has never done anything unworthy of trust.” On revenue sharing, 唐凯 said “the time has not yet come”; for now, the priority is to make the pie bigger together.

8. The automobile logic: take full responsibility for the outcome, allocate by value contribution

• 任向飞 highlighted the difference in accountability between cars and phones. If a phone service fails, the app provider is responsible; if Alipay has a problem, nobody blames OPPO. But the automaker is responsible for every problem involving the car. His vision is for the car, as “highly tangible embodied intelligence,” to detect a driver’s physical distress and activate autonomous driving to take the person to the nearest hospital. The 阿福 foundation model could diagnose the situation in advance and book the hospital and relevant department. This is chain-based ecosystem coordination, not a business model confined to a single ecosystem position.
• The auto industry was “relatively behind” during the traffic and information monetization era and generally did not charge users for services. The next phase shifts from information and traffic to a service-and-delivery economy: first create value that makes users willing to pay for end-to-end delivery, then determine the split. “Allocate according to value contribution: whoever provides the core data, the key central decisions and the final delivery will receive the best pricing from the market.”

9. The 10M-scale Agent economy: the 30,000-Skill debt and next year’s second-half inflection point

• 任向飞’s vision is that once the market reaches 10M scale, “many more Agents providing atomic-level services will appear in large numbers,” bringing a wave of individual and small-team developers and potentially an “Agent internet era.” Terminal integration and primary-Agent capability will become decisive. Geely plans to develop a dedicated AI OS centered on task delivery and capable of calling surrounding Agents.
• 周围 pushed back with the roundtable’s coldest assessment: “walking on thin ice.” Washing a car and booking an appointment may look the same as before, but “the essential difference is not the service itself; it is what has changed in the machinery supporting that service”—the foundation model plus Harness. Harness marks a huge shift from last year on PC precisely because it has fully adapted to more than 30,000 Skills across the system. On phones, that work “has only just begun… not one of those 30,000 can be missing,” and completing it will take “at least a year.”
• 吴嘉 was relatively optimistic: extreme personalization, lower barriers and a massive expansion in supply are arriving simultaneously, creating a very large opportunity. China’s Agent supply “does not need to be worried about at all”; anyone with a specialized skill, individual or organization, can build an Agent. “They just don’t know where to go to provide the service”—the problem Alipay’s Connect conference is designed to solve. 唐凯 offered a different angle: mobile internet is fundamentally an attention economy, while mature Agents will give people more time. “People should be exercising and focusing on health,” making Ant’s 阿福 “very forward-looking.”
• 韩歆毅’s closing view: industry organizational capacity may never catch up with model capability, but the pace of technological change “could be faster than any of us imagine.” The key to acceleration is building a commercial loop that lets every participant see the returns. “Let me be bold”: some merchants may feel the initial value within 3 months; by around 6 months, roughly after next year’s Spring Festival, more merchants may join; the inflection point may come in the second half of next year, and the true breakout could arrive in the second half of next year as well.