China’s Personal AI Agents Stall on Ecosystem Access, Not Model Power

NextFin News — Personal agents were supposed to arrive with the holiday season. In late September, reports described Doubao teams racing to produce an internal experience build within days, while ByteDance moved to tighten integration between a project code-named Spell and the main Doubao conversation product. Alibaba’s Qwen App lead publicly committed to speeding up a Personal Agent. Tencent’s Yuanbao was also said to be preparing a version, though timing stayed vague. The expectation was straightforward: once domestic versions matched the continuity shown by Meta’s Muse, users could hand off multi-step tasks without opening every App themselves.

That expectation has already met resistance. Meta’s Muse impressed with long-running, cross-service workflows—pulling a school form from email, checking a calendar, and seeking confirmation before sending. Shortly afterward, Amazon restricted the agent’s access to its store, citing lack of prior notice, undisclosed automated access, and concerns over credential handling. The episode underscored a basic split: user consent does not automatically equal platform permission.

Chinese companies face the same split in denser form. Daily life is fragmented across Apps that control accounts, catalogs, payments and fulfillment. An agent that can read a screen or hold a few API keys still needs each service provider to accept the traffic, define allowed actions, and settle revenue. Mis-clicks, disputed orders and privacy rules land on the merchant or platform that actually fulfills the request. When comparison and decision-making shift upstream into an assistant, the destination platform can lose impressions, recommendation slots and incremental discovery.

The result is a dual identity problem. As builders of personal agents, the large platforms want external services to open wider. As operators of core businesses, they must weigh whether someone else’s agent will interpose itself between them and their users. Everyone wants to be the default entry point; few want to become a silent backend for a rival’s assistant.

Putting the agent on the phone does not dissolve the constraint. The consumer version of Doubao’s phone assistant, launched with Nubia’s NaviX Ultra, demonstrated cross-App flows such as extracting event details from a photo into a calendar or summarizing Feishu messages. Yet ByteDance also published a Screen Automation Execution Protocol that lets third-party Apps opt in or restrict screenshots, simulated input and content changes. In independent tests, the assistant could complete a coffee order inside Douyin after authorization, but could not drive Taobao Flash Purchase, JD.com or WeChat messaging. Success depended on which platform hosted the final step.

ByteDance has responded by recruiting partners. Caocao was among the first ride-hailing services linked to the phone assistant. Ahead of the National Day period, Doubao concentrated map, mobility and ticketing functions into a prominent “Travel with Doubao” surface, drawing maps from Baidu, rides from Caocao, and tickets from specialized providers, while hotels often routed into Douyin’s inventory. Each expansion still requires interface work, authorization design and ongoing maintenance beyond the model itself.

Alibaba is pursuing a different division of labor. At the Yunqi conference it offered Qwen Intelligence as a modular suite of planning and cross-App execution capabilities that handset makers can adopt in whole or in part. Honor became the first partner, shipping related agent features on the Magic9 series. Honor’s own YOYO assistant already extracts times and locations from tickets or calls, creates to-dos, and surfaces Alipay service cards for limited actions. The pattern is clear: the model provider supplies reasoning and orchestration; the phone maker keeps the system entry point; the service platform decides which steps remain user-confirmed.

Internal ecosystems lower the initial cost. Alibaba can push Taobao, Tmall, Taobao Flash Purchase, Amap and Fliggy toward Qwen under group coordination. ByteDance can reuse Douyin e-commerce and local-life supply, and has already recorded measurable hotel-order growth attributed to Doubao. Tencent can start from WeChat’s native messaging, payments and mini-program surface, returning cards that hand the user into a developer’s page for final selection and payment. In each case the company avoids rebuilding merchant acquisition and transaction plumbing from zero.

Those advantages are real and uneven. Alibaba covers more everyday verticals inside one corporate boundary; ByteDance must still court external map, mobility and ticket providers to match breadth; Tencent inherits a vast mini-program network but still needs developers to expose agent-friendly interfaces and accept new economics. Even internal reuse is not free. Inference, task-state management and error recovery consume resources. New channels can raise merchant fees—Doubao hotel orders, for example, have been assigned a distinct higher rate than ordinary Douyin traffic—so merchants will continue only if the assistant delivers incremental demand.

Capital budgets already run into the tens of billions of yuan per quarter across the major groups. Personal agents sit on top of that spending. They do not escape the requirement to share margin with whoever owns inventory and fulfillment. Analysts have already asked whether an agent that merely relocates existing mini-program transactions will raise compute cost while reducing ad exposure. Executives answer that better experience should expand the ecosystem and that existing monetization can still apply if costs are controlled. The answer remains to be proven at scale.

Six days may be enough to ship an internal demo. Winning daily use requires the agent to complete the tasks users actually have, across the Apps they already rely on. Domestic companies can move fastest where they control both the assistant and the service. Everywhere else they must negotiate access, identity, liability and revenue. Model quality is no longer the binding constraint. Platform consent is. Until that consent broadens, the seamless personal agent remains more aspiration than default infrastructure.