China's largest consumer platforms are converging on the same product: an always-on assistant that holds a long-term goal and works on it in the background, rather than answering one question at a time. Tencent is testing a personal agent internally codenamed Handy Bot and plans a standalone app, with a service account already live inside WeChat describing itself as "Your Personal AI Agent." ByteDance's Doubao personal-assistant project, in internal development since April under the codename Spell, has been named Xiaodou and is set for an independent app. Alibaba is repositioning Qwen as "My Qwen," built around personal context and memory.
The trigger was Meta's Muse, launched on September 8. Sensor Tower estimated roughly 730,000 downloads in the first five days, a first-place finish on the US App Store free chart by day ten, and about 2.6 million downloads across iOS and Android by day thirteen. Meta's shares rose 11.43% in a single session at the peak of the attention, adding close to $200 billion in market value. OpenAI answered with Dots at its September 29 DevDay, and Google's Gemini Spark sits in the same category.
What separates a personal agent from a chatbot is where the task lives. A chatbot ends the task when the conversation ends. A personal agent keeps running after the user closes the app: watching a fare, answering mail, comparing prices, chasing a refund. Meta's design gives each user a dedicated Linux virtual machine, keeps the agent and the connected data inside it, and places a separate gatekeeper agent, Sentinel, in front of any outbound action; credentials sit in separate secure storage that the agent can use but not read. Spending and sending stop for confirmation.
Tencent's approach is spread across three parallel efforts, which is characteristic of the company. The first is Xiaowei, a native WeChat assistant in gray testing since June, reachable by text or voice to operate WeChat's own functions — sending messages, placing calls, changing settings — and to invoke mini programs, such as ordering a coffee. The second is LightVela from Tencent Cloud, a hosted agent instance that runs around the clock, holds the user's data, and appears as a contact inside WeChat, QQ and WeCom after a QR-code authorization. Its origin is telling: during the OpenClaw wave in March, Tencent ran a free installation event that drew more than 70,000 people, and concluded that what users wanted was not a one-time setup but a persistent, maintenance-free AI service. The third is Handy Bot itself, positioned as a single entry point that unifies the company's agent portfolio, spanning Yuanbao, QClaw, ima, WorkBuddy and CodeBuddy, and can dispatch mini programs across personal and work contexts.
ByteDance is growing its agent from the phone outward. Doubao's consumer phone assistant shipped on Nubia's NaviX Ultra in September with the ability to act across apps from a single sentence, and Spell is led by the Doubao phone-assistant team — in effect lifting the "operate the phone for you" capability out of the handset and into a standalone product. In August, ByteDance folded its TRAE and Coze teams into Doubao and launched Doubao Work, consolidating agent capability under one brand with parallel consumer and enterprise lines.
Alibaba's bet is on memory and time. At its Apsara Conference on September 22, Qwen said it would accelerate a Personal Agent build-out for 300 million users. Product lead Zheng Sishou framed the distinction as knowing not only what happened but what is persisting: one wrong answer means little, while the same knowledge point missed repeatedly is a signal; one bad night is random, while a pattern across months is a health trend. With user authorization, Qwen connects health and fitness data, professional financial data and, for some users, personal holdings; its open platform now covers more than 20 domains. The company has held back on taking initiative, stepping in only where intent is clear — for example, offering to set a monitor on a stock a user has been following.
None of this is settled by model capability. A personal agent that genuinely handles daily affairs has to connect to shopping, transport, email and payment services, and the platforms that own those relationships may be reluctant to hand the user and the transaction entry point to someone else's agent. That makes trust and ecosystem access — not benchmark scores — the binding constraint. It also explains the concentration of the race: the companies pushing hardest are the ones that already own a messaging surface, a handset or a payment rail, because the entry point, not the model, is what the customer relationship hangs on.
Comments (0)
Log in to join the discussion
Log InNo comments yet