Article title: "I'm Not That Optimistic About Big Tech Building Personal Agents"
Body: "Among the major Chinese internet companies, which one are you most optimistic about building a Personal Agent?"
Over the past month, ever since the two products Instinct and Muse went viral in North America, investor 刘墨昱 has had this topic brought up repeatedly every time they meet with peers and friends.
Yet this happens to be a question with no consensus. 刘墨昱 and friends carefully dug into each major internet company, and each one appears to hold a massive ecosystem foundation, but when you really strip it down, each company's obstacles to building a Personal Agent are just as numerous.
"From all the discussions so far, I'm still not that optimistic about major Chinese internet companies doing Personal Agents well," 刘墨昱 shook their head.
Which team should Personal Agent be handed to?
The first hard problem for big tech building a Personal Agent sits right inside the organization: who exactly should be put in charge of this?
Inside ByteDance, they chose to hand this to the Doubao phone assistant team to lead.
In April this year, a project codenamed "Spell" was already in internal testing. In the eyes of many industry practitioners, the phone assistant team had long been doing system-level cross-app operations—that is, letting AI jump between multiple apps for users, compare prices, and place orders. This is basically the same thing a Personal Agent needs to do, so ByteDance didn't need to build a team from scratch.
But curiously, there were reports early on that after products like Muse became hugely popular overseas, the project was accelerated, yet to this day it still hasn't been released externally. Some industry insiders speculate this may be related to the team's ongoing independent spin-off financing, as the spin-off process involves many uncertainties that would change product development progress accordingly.
At Alibaba, they handed the Personal Agent project to the Qwen team, with Wu Jia in charge. This also makes sense—Alibaba established the ATH business group in March to integrate its core AI segments, with Qwen responsible for AI to C and DingTalk responsible for AI to B. Handing Personal Agent to the Qwen team follows naturally.
On October 13, the Qwen team also took the lead in launching Qwen Go. But notably, they didn't release Qwen Go as a standalone app—they embedded it inside the Qwen app.
Some practitioners believe the advantage of this approach is that Qwen is already recognized as Alibaba's main AI to C entry point, with Alibaba disclosing its C-end monthly active users exceed 300 million. Qwen Go evolves on this foundation without needing to acquire users from scratch.
But product person 刘凡宇 says there are also concerns: "The chatbot DNA and the personal agent form aren't entirely consistent." When users open Qwen, their expectation is to "ask questions"; but a personal agent wants to get users accustomed to "assign a task and walk away." Putting two different expectations into the same entry point creates tension in product design.
Tencent's situation is the most peculiar—who should be the main force building the Personal Agent seems undecided.
The Yuanbao team is working on it. In October there was market news that they were pushing forward "Yuanbao Assistant" and planned to launch a standalone app. The team's head is Wu Zurong, who previously created the hit product Tencent Meeting. WeChat is also working on it—they've already incubated their own personal agent "Xiaowei."
"Of all the companies and teams, I'm very optimistic about WeChat doing this," 刘凡宇 believes. WeChat possesses all the raw materials a Personal Agent needs—user scale, services, payments, personal context—but right now WeChat happens to be the most cautious about doing this.
Two people close to Tencent's senior leadership thinking say that internally Tencent has always been very cautious about integrating AI into WeChat, unwilling to let "half-finished products" disrupt WeChat's sense of stability. Currently Xiaowei is also an exclusive assistant within the WeChat ecosystem, not touching the external internet.
The Unavoidable Internal Resource "Battle"
Deciding who does it is just getting the cards.
To actually build the Personal Agent product, you still need compute, traffic, and interfaces. The group's pool is only so big—when you get more resources, other teams get less. This will likely trigger an internal resource game.
Who will allocate your compute resources?
The compute math for a Personal Agent is not the same as for a chat product. Chat is a user asks one question, the model answers one question, with costs calculated per call. But a personal agent needs to reside in the background, remember the user, and run long tasks—equivalent to giving each person a "dedicated computer," with costs calculated per person, per day.
After OpenClaw went viral, a team that once tried to productize a personal agent product shared the costs of building and running the product at the time: equipping each agent with an independent computing environment and a reliable network identity requires at least about $50 per user per month (excluding tokens).
Put this math inside big tech, and it becomes a resource problem. Compute is a unified pool at the group level, and no one has plenty to spare—especially this year with the AI Coding explosion and accelerating multimodal commercialization.
An informed source once mentioned this detail to Leiphone: last year Tencent was extremely short on compute and rented quite a few GPUs from another big tech company, but this year GPU resources became even scarcer, so the big tech company had to take back all the GPUs previously rented to Tencent and even paid Tencent a penalty fee.
This is the current state of compute usage among vendors—resources are extremely tight. If a personal agent takes one more piece, someone else gets one less.
Who will get less? At least three types of "long-term residents": first, model training—Seed and Hunyuan are both catching up; second, office agents—Tencent's WorkBuddy, Alibaba's Qwen Office, ByteDance's Doubao Work—whose payment paths are relatively clear; third, the original chat products, which have large user scale and high call volumes.
A Tencent Hunyuan colleague revealed to Leiphone (public account: Leiphone) that a few months ago, when WorkBuddy was hugely popular, Tencent increased investment and various BGs diverted compute to support it, which also affected multimodal and pretraining to some extent.
Who will give you interfaces?
The Personal Agent model only determines how smart the product is, but the ecosystem determines the product's life or death.
When the Agent needs to connect to Taobao, WeChat Pay, mini-programs and other businesses, whether the business lines cooperate will be another major challenge facing the Personal Agent team.
Alibaba is the smoothest of the three. In January this year, Qwen integrated Taobao, Alipay, Fliggy, Amap and other businesses, launching over 400 service functions; after authorization, Qwen Go can also call data from Taobao, Amap, Damai, and Quark Netdisk.
But behind the smoothness there's a subtle arrangement: business lines like Taotian and Amap are simultaneously AI-ifying their own services, building vertical Agents in their respective domains, and Qwen's relationship with them is compared to "central and local." Business lines must both open interfaces to Qwen and defend their own Agents and entry points—contradictions may emerge in this process.
Tencent's interfaces are held in WeChat's hands, and WeChat's approach is restrained. For Xiaowei to call mini-programs, it relies on integration guidelines published by WeChat developers, with developers choosing whether to integrate; when calling operations like sending messages and transferring money, each step still requires user confirmation.
Within Tencent itself, if Yuanbao wants to connect with WeChat, that's even harder. Do you all still remember—during the 2026 Spring Festival, Yuanbao once launched a product called Yuanbao Pai, and at the time they aggressively gave out red envelopes for viral growth during the holiday, but before three days of promotion, the link was blocked by WeChat.
History Always Repeats Itself
Resource competition between internal departments also has historical precedent.
Around 2012 to 2015, when the mobile internet wave swept through major internet companies, similar dramas already played out. At the time, each company's PC business and mobile business were placed in different departments:
Baidu: The Search Business Group (SSG) was led by Xiang Hailong, and the Mobile Service Business Group (MSG) was led by Li Mingyuan; Tencent: PC QQ, Qzone and other social businesses belonged to the Social Network Group (SNG) led by Dowson Tong, while mobile QQ remained in the Mobile Internet Group (MIG) led by Liu Chengmin; Alibaba: Taobao's PC and wireless were two business lines—Taobao.com (PC) belonged to the Taobao Business Group; mobile Taobao belonged to the Taobao Wireless Business Unit.
Baidu: The Search Business Group (SSG) was led by Xiang Hailong, and the Mobile Service Business Group (MSG) was led by Li Mingyuan;
Tencent: PC QQ, Qzone and other social businesses belonged to the Social Network Group (SNG) led by Dowson Tong, while mobile QQ remained in the Mobile Internet Group (MIG) led by Liu Chengmin;
Alibaba: Taobao's PC and wireless were two business lines—Taobao.com (PC) belonged to the Taobao Business Group; mobile Taobao belonged to the Taobao Wireless Business Unit.
The result of this PC-mobile separate governance was that old and new departments fell into extremely fierce internal friction and wrangling over resources and entry points, and new mobile businesses developed very slowly.
Take Taobao: by the end of 2013, Wu Yongming had taken over mobile Taobao for over a year, and DAU got stuck at just over 40 million. At the time, senior leadership began to worry, feeling that "Wu Ma" couldn't lead it, and by the 2026 Spring Festival, the Alibaba Group decided to directly replace the person, bringing in 樊路远, who had done well with Alipay DAU. An informed source said: "Wu Ma, completely unaware at the time, 'turned green with anger,' but the die was cast." In reality, 樊路远 lasted less than three months before leaving.
Baidu was the same—originally in 2012-2013, Li Mingyuan had already led mobile Baidu to over 100 million daily active users. At the time, PC-side Xiang Hailong's job level was one grade higher than Li Mingyuan's, and in specific traffic allocation, R&D resource investment, and strategic priority, the mobile team could never outcompete the PC team. Mobile Baidu, which had already seized the initiative, had its development speed severely slowed.
The most regrettable was Baidu. The mobile side was suppressed by the PC side for years, until 2016 when Li Mingyuan was reported for economic issues and ousted, and this internal friction finally came to an end—but the golden window for mobile had long been missed.
History doesn't repeat, but it always rhymes.
Now big tech is pulling out the teams working on large models and Agents and setting them up as "innovation departments" at the same level as core business departments (like e-commerce, local life). Once this architecture appears, the "end-side battle" of the past immediately becomes today's "battle between AI teams and traditional business teams."
Everyone on the surface is working hard for the company's future, but in reality each carries different KPIs: old departments must protect revenue and profit, new departments must push AI and grab traffic—and in the end it's still wrangling, setting up roadblocks, hoarding data and interfaces tightly, and no one can move anyone else.
If these organizational conflicts are handled improperly, these vendors will likely find it hard to build Personal Agents well.
These contradictions ultimately land before the number-one position: when old businesses must protect revenue and new businesses must grab the future, which side does the company put its resources on?
How does the number-one position balance the future and the present?
For vendors like Alibaba, Meituan, and Baidu that excel at "connecting services and transactions," the contradiction is especially direct: they must build Personal Agents, yet may not dare to fully let go and do it.
The reason they must do it is simple—not advancing means falling behind. If they don't go all-in on Personal Agents, and a startup or competitor succeeds in building a super AI assistant that takes away all consumers, then they'll face having the rug pulled out from under them and being eliminated.
BAT Building Personal Agents: