Yuanchuan Research Institute: Following the Personal Agent Trend Is Not Advisable

The annoying AI-office ads are still rolling on elevator screens, and the big tech companies have set their sights on another big trend: the Personal Agent.

Alibaba's Qwen assistant has been upgraded to Qwen GO, which can break away from the Alibaba ecosystem and autonomously operate various authorized platforms to complete complex, long-horizon tasks.

ByteDance is reportedly also hurrying to develop a personal agent, codenamed "Spell," a continuation of the Doubao phone assistant in the Personal Agent direction—my apologies, so it turns out to be the reincarnation of the Doubao phone.

Tencent hasn't been idle either. From the end of September to now, multiple Personal Agent products have been exposed internally one after another, and a new round of horse racing has already begun.

The chatbot battle during Spring Festival is still vivid in our minds; this time there were no red envelopes to attract new users, which is progress.

Where is the wind coming from?

Domestic big tech companies are all following the trend en masse, and you could say the responsibility all lies with the American side.

On September 8, Meta launched its personal agent product Muse, limited to the US and Canadian markets, but it quickly topped the App Store free chart after release, surpassing ChatGPT's short-term data from back then, with downloads exceeding 10 million within 40 days of launch [1].

Among the big American AI companies, Meta took quite a few detours and had always maintained a relatively low profile. But with Muse's sudden emergence, Meta went from being the American Baidu to a leader in Personal Agents in one stroke, with market cap gains exceeding $200 billion (it later fell back), and who wouldn't be envious seeing that.

English text is being prepared.

Meta also places great importance on Muse, on one hand leveraging the traffic advantages of its own platforms like Facebook and Instagram for high-intensity advertising bombardment; Chief AI Officer Alexandr Wang, hired for $14.3 billion, also personally stepped in, spamming at high intensity on the competing platform X.

Muse also incidentally boosted a large number of startups in the same track, such as Noah Shinn, who was recently hailed as the "hottest post-00s" figure; his company Instinct just raised funding at a $2.5 billion valuation at the end of August, and by mid-September its valuation had shot up to $10 billion.

Instinct is like a real human assistant in your contacts whom you can reach at any time. Users don't need to download a dedicated app; they communicate with users via SMS, iMessage, WhatsApp, or even phone calls.

Users can, just like messaging a friend, directly instruct it to book restaurants, arrange travel, call customer service, organize emails, or even contact merchants to resolve issues. It can also remember users' preferences, continuously handle tasks, and proactively report progress—it likewise has a very good eye for work.

Regarding this situation, Alexandr Wang's assessment was:

Muse completely crushes Insect (Instinct)

Three weeks after Muse's release, OpenAI also swiftly launched its personal agent Dots, focused on all-weather task execution and integrated into personal and team workflows.

Overall, compared with previous apps for shopping, ride-hailing, and ordering milk tea, Personal Agents can connect to context hubs such as email, SMS, and calendars to understand users' context, thereby handling more complex tasks. The tool attribute is weaker, and the assistant attribute is stronger.

Not acclimated to the local environment

On the day Muse launched, Zuckerberg discussed commercialization in an interview: short term relies on user subscriptions, long term relies on transaction commissions [2].

First, subscriptions. On the eve of SpaceX's IPO, Musk did some calculations in the prospectus: in the total AI market size, infrastructure is $2.4 trillion, enterprise applications $22.7 trillion, and as for the scale of C-end subscriptions and advertising, even the bold and daring Mr. Ma only estimated $1.36 trillion [3].

At that time, the concept of Personal Agents was not yet this hot, but in the programming field, which already has a closed commercial loop, the current situation is not very optimistic either.

In October this year, SemiAnalysis published a report calculating Anthropic's revenue structure: the main force is API customers charged by usage, while Claude subscription users consumed 41.7% of compute power but contributed only 10% of revenue [4].

Compute consumption is about the same, but spending differs by nearly ten times. Both are buying tokens, yet when paying with one's own card people are meticulous, and when asking finance for an invoice they are lavish.

Compared with monthly subscription fees, participating in users' daily consumption and taking commissions has greater commercial imagination space. But this model seems not quite suitable for the constitution of Chinese babies.

An Agent handling various affairs for users needs to call existing internet products. On this matter, there is no small difference between the Chinese and American internet ecosystems.

American internet services are relatively dispersed; shopping and booking hotels sometimes require going to official websites, and users need to open different websites, register accounts, fill in addresses, enter credit card information, and complete payments—the operations are extremely cumbersome, giving browser-based Agents room to be useful.

China's mobile internet ecosystem is extremely prosperous, even a bit pathological: users can browse Xiaohongshu in Dianping, watch short videos in Alipay, book hotels in Douyin, watch short dramas in Meituan, and take out a loan in Amap to buy a car and register as a ride-hailing driver.

Consumers' high-frequency needs can all be fulfilled within super apps. Originally, it took just a couple of taps to get things done, but now they have to explain their needs to an AI, grant permissions, wait for execution, and finally approve the results. Against super apps, AI may not be competitive.

At the same time, the philosophy of the Personal Agent also directly collides with the business model of the mobile internet.

The premise for monetization in the vast majority of mobile internet business formats is that users stay and interact within the product as frequently and for as long as possible, monetizing users through advertising, e-commerce, and other means. So "app usage time" is critical—Attention Is All I Need.

Personal Agents need internet products to provide services, but the time they save users is precisely the foundation of internet product monetization. The traffic gateways, advertising exposure opportunities, and some transaction revenues that originally belonged to these platforms may all be intercepted by Agents.

In China's internet, where super apps are highly concentrated, it will only be harder for an AI assistant to cross the walled gardens of each platform.

In October this year, one month after the consumer version of Doubao Phone launched, it issued a tearful statement saying it had been blocked by 151 apps. How could Doubao be allowed to sleep soundly beside the emperor's bed?

You should know that less than half a month after Muse launched, it had already dragged down the stock prices of hotel booking platforms like Expedia. At the time, Bloomberg, never one to miss a spectacle, proposed a hypothesis: if Personal Agents diverted 5%–10% of business from travel, ride-hailing, and food delivery platforms, the companies involved could lose more than $5 billion in revenue [5].

Translated to China, business departments carrying OKRs cannot bear such a cost, big companies' management cannot bear it, and deep-value investors in the Hang Seng Tech sector certainly cannot bear it either.

Entrance Anxiety

According to Sensor Tower, Muse's day-one retention rate after launch was as high as 73.3%, almost twice that of ChatGPT [6].

But the novelty faded just as quickly: by day 14, Muse's retention rate had dropped to 22.5%, surpassed by ChatGPT's 26.8%.

Yet Agents are also an expensive business. Every user served by a large model incurs a corresponding cost.

Morgan Stanley estimates that, depending on token consumption, Muse's monthly cost to serve a single user ranges from $3 to $128 [7].

There is a big backdrop to Meta doing Muse: Meta's large models are too weak, so the compute it purchased had to be rented out, and compute was not that scarce. Add to that Meta's size and resources, and it can afford long-term investment.

Compared with the United States, China's compute resources are even scarcer, but that cannot stop big tech's desire to advance.

The vision of a Personal Agent coordinating all apps is too tempting. Companies absent in advance may face the risk of their original business models being restructured. Although the market space is unclear, costs can be calculated, but the risk of missing out cannot be measured. If life or death is uncertain, then it is alive.

Everyone wants to block Doubao Phone, and everyone wants to become Doubao Phone. If every company allows Agents to operate only within its own ecosystem, then what consumers ultimately get may be a group of digital assistants, each ruling its own hill and unable to communicate with one another.

Microsoft Research proposed a hypothesis in "The Agentic Economy": even if Agents already have the technical ability to communicate with one another, commercial interests may still push platforms to build closed Agent ecosystems—so-called "Agentic Walled Gardens" [8].

In this way, the problem of fragmented internet services that Personal Agents were originally meant to solve would instead reappear in another form.

Epilogue

Personal Agents provide a new narrative for the fragile sentiment of the capital market.

The industry generally believes that if Coding were the only trillion-dollar-scale market, the current level of capital expenditure would already be sufficient.

But the theoretical market space for Personal Agents is much larger than Coding, and naturally consumes more compute, so compute and storage keep the music playing and keep dancing. Deploying Agents in the cloud will in turn drive demand for CPUs, and AMD has already happily pocketed that part of the valuation.

But no matter how good the shovel business is, someone still has to dig up gold. Acquiring a bunch of users and consuming a bunch of compute without lifting Hang Seng Tech—then it really is a huge loss.

References

[3] Space Exploration Technologies Corp. — IPO Registration Statement, SpaceX

[8] The Agentic Economy, Microsoft Research

This article comes from the WeChat public account "Yuanchuan Institute" (ID: YuanChuanInstitution), author: Zhang Beijia, published by 36Kr with authorization.