Back in April 2016, Mark Zuckerberg stood on stage at the F8 developer conference in San Francisco, demoing chatbots inside Messenger. He sent a message to the bot belonging to flower retailer 1-800-Flowers and ordered a bouquet.

Standing before the audience, he declared: "I've never met anyone who enjoys calling businesses." He then added that ordering flowers would no longer require a phone call to the floral company in the future.

At the time, Messenger boasted over 900 million monthly active users and had partnered with more than 40 companies, including CNN, eBay, and Walmart. Zuckerberg's logic was simple: nobody wants to install a separate app for every merchant, and the chat window could replace them all.

On September 8, 2026, the same company unveiled its personal AI agent product, Muse. Users send messages to it via a standalone app or through WhatsApp, and the agent contacts merchants and completes purchases on their behalf. The underlying logic bears a striking resemblance to what was pitched a decade prior.

A lot has changed over those ten years, with massive strides in technology. Muse operates within individual virtual machines allocated per user in Meta's cloud, capable of controlling a browser, filling out forms, comparing prices, and handling emails. It continues executing tasks in the background even after the user closes the app.

Within just 13 days of launch, Muse climbed to the top of both the US iOS and Android free app charts, racking up over 2.5 million downloads.

For Meta's stock price and its narrative, the past decade has traced a full circle.

The "conversational commerce" vision of 2016 never became the gateway it was hyped to be. Messenger bots ultimately devolved into customer service tools, and the human-backed assistant, M, was shut down in early 2018.

In 2023, Zuckerberg proclaimed that year as the "year of efficiency," initiating layoffs, flattening management layers, and lowering capital expenditure forecasts. Meta's shares rose nearly 200% over the entire year. Back then, the market rewarded cost-cutting.

By 2026, Meta has raised its full-year capital expenditure guidance to at least $130 billion. Its second-quarter free cash flow dwindled to just $784 million. On September 21, the 13th day of Muse's launch, Meta's stock surged 11.4% in a single day. This time around, the market is rewarding spending.

Investors have been waiting for a consumer-facing use case that could justify the astronomical capital outlays, and Muse arrived precisely at that moment.

Every leap in AI technology prompts tech giants to attempt claiming a super gateway for consumers. The power of this imagination has always far outweighed its reality.

But what determines whether something can become a gateway is never just whether it can get things done. It also hinges on who pays after the task is finished, who is willing to grant access, and whether users will stick around.

Muse has answered the first question better than any of its predecessors, but the remaining three questions remain as unresolved today as they were ten years ago.

Attempt Number Five at Building a Gateway

Every two to three years, the narrative of an "AI-driven new gateway" returns with greater technical sophistication and even more market enthusiasm.

When Amazon released the Echo in 2014, Alexa was expected to become the hub of home consumption. Yet by 2018, only about 2% of Alexa device users had placed an order via voice, and approximately 90% of those never tried it a second time. While Amazon disputed these figures, which were attributed to sources cited by The Information, it offered no alternative data.

In the two years following Messenger's bot announcement at F8, its presence steadily faded, and Facebook stopped treating Messenger as the primary battlefield for its commercial platform.

When ChatGPT first captured the public's imagination in 2023, the GPT Store and plugins were touted as the "App Store of the AI era." Around the same time, Rabbit R1 and Humane AI Pin launched with the narrative of "replacing the smartphone." In March 2024, A-shares in China saw a wave of limit-up trading on Kimi concept stocks, with Kimi's labels at the time being long-context processing and a 2C consumer focus.

The outcomes were underwhelming: GPT Store traffic was lackluster; Humane sold its assets to HP in early 2025, and the AI Pin was promptly discontinued; Kimi's subsequent growth primarily came from overseas API calls, and the story of it becoming a consumer-facing gateway never materialized.

On September 29, 2025, ChatGPT introduced Instant Checkout, allowing users to complete purchases directly within a conversation. On the announcement day, Etsy's stock jumped nearly 16%. Within less than six months, OpenAI scaled back this feature, handing checkout responsibilities back to merchants. Walmart executive Daniel Danker disclosed that conversion rates for in-app checkout were only one-third of those achieved when redirecting users to Walmart's website.

More recently, during this year's Spring Festival, Alibaba's Qwen used a 3 billion yuan free-purchase campaign to complete 120 million orders in six days. QuestMobile data showed its DAU briefly surged to 73.52 million, though it settled around 30 million after the peak.

Muse now marks the fifth attempt. The common thread across all of them is that the market has paid for the hype almost every time, yet each initiative has ultimately stalled at being a "feature" rather than evolving into a "gateway."

Failures are often blamed on technology. The truth is, speakers from over a decade ago could already understand "buy another pack of tissues," yet users tried them once and never returned.

Muse's experience is certainly a significant step forward. Morgan Stanley analyst Greg tested Muse to buy a hat, and from sending the request to receiving the confirmation email took less than two minutes. Muse even proactively reminded him about his flight time and suggested expedited shipping. The experience Zuckerberg promised a decade ago has finally been achieved.

However, many of the barriers are not about the models. On the day after Muse's launch, payments media outlet PYMNTS tasked Muse with restocking on Amazon, ordering pizza from Domino's, and making a reservation at Resy. All three tasks failed, blocked by account authorization, the need to connect Gmail and calendars individually, and the checkout process.

If the gateway story fails again this time, the reason may still have nothing to do with technology.

Time: To Be Wasted or Eliminated

The market's belief that personal agents could become a new gateway largely stems from the sensory experience of the past decade, where feed-based recommendations reshaped the gateway landscape of the internet.

Douyin's algorithm-driven recommendations have wrested significant user time away from search and social platforms. Citigroup, citing QuestMobile data, noted that in July this year, Douyin's daily active users reached 714 million, with each user spending about two hours per day on the platform, and total usage time surpassing WeChat for the first time.

Short videos, live streaming, and feed ads are all fundamentally designed to extend user dwell time, manufacturing demand within that extended stay.

But the success of recommendation feeds is premised on people having vast amounts of time they are willing to spend.

For many users, scrolling through Douyin or browsing Pinduoduo is entertainment in itself. Colin Huang once described Pinduoduo in a letter to shareholders as a combination of Costco and Disney, where the Costco half represents efficiency, and the Disney half represents killing time.

The product logic of a personal agent is precisely the opposite.

Its value lies in eliminating time. The user hands a task to the agent, which completes it in the background, freeing the person to do something else.

Muse pushes this to the extreme. Users can close the app, and it keeps working, only returning to the human for approval when necessary.

Personal agents might generate demand. Muse can turn recipe Reels saved by users into shopping lists, suggest menus for dinner parties, remember friends' dietary restrictions, and proactively send out invitations. Qwen's agent, opened to brands, is designed to provide itinerary reminders, membership expiration alerts, and repurchase recommendations.

The Morgan Stanley analyst also observed that Muse proactively tells users "I can do this for you." Silicon-based assistants' recommendations and reminders can sometimes prompt people to buy extra items.

The demand it generates has two characteristics.

First, it is highly dependent on pre-existing intent or records. Subscriptions renewals, repurchases, exchanges, or switching to cheaper insurance are mostly replacement demand, rarely created from thin air.

Second, it does not generate dwell time. Recommendation feeds retain people when they have no specific purpose, and every extra minute of stay equates to one more impression. An agent's entire effort is to get the user to leave as quickly as possible, finishing the task and moving on.

A reviewer from MBI Deep Dives quickly burned through 81% of the free tier's weekly quota after Muse launched, believing its best use was connecting all his Gmail accounts. As for ordering food, he wrote that at least half the time he doesn't know what he wants to eat, and the decision only forms while browsing DoorDash, the American equivalent of Ele.me.

For those who already know what they want, an agent improves efficiency. For those who haven't figured it out yet, browsing is the decision itself. Skipping the browse means skipping the phase where demand is formed.

Just as Muse can turn recipe Reels into shopping lists, the initial thought to cook arises within Instagram's recommendation feed. The agent merely takes over the execution.

No matter how fervent the AI evangelism, one shouldn't credit Instagram's recommendation feed's contribution to the agent simply because AI is newer and more exciting.

Once an agent starts proactively recommending, it drifts toward a recommendation system. And in e-commerce and local services, revenue from recommendation slots naturally comes from the seller side.

In early June, a media outlet tested Doubao's price comparison for the same pair of headphones across three platforms. At the bottom of the answer, however, was a product card linking to Douyin's own mall.

Accepting money from one side means advocating for that side. This is the fundamental dilemma agents face.

Orders Without Visibility into the User

Search engines have established one of the most successful business models on the internet.

Its core exchange is simple: users get information for free, search engines capture intent, and merchants pay for that intent.

But search itself is also a tool for eliminating time. The faster a user finds an answer, the better. It becomes a gateway through delivery, completely unrelated to dwell time.

A search engine delivers a person with intent to a merchant's website. Once there, everything—browsing, adding to cart, repurchasing, member conversion, and brand impression—belongs to the merchant. The merchant pays for this referral because a single click can turn into a long-term customer.

Monetization of internet gateways largely follows these two paths: either waste users' time like recommendation feeds, or hand users over cleanly like search, carrying them in and letting them go.

Agents sit on neither side. They eliminate time but want to keep users within their own grasp.

What they hand to the merchant is an order. The merchant sees the account, shipping address, and the transaction, but cannot see the person who was browsing the store. There is no browsing history, no dwell time, no attraction to other products on the homepage. Nobody remembers the brand.

Next time a price comparison occurs, the agent will start from scratch. Winning this order does not constitute an advantage for the next one.

For merchants, agent-mediated transactions are closer to one-off deals. This difference will inevitably influence merchants' willingness to pay.

Search can be free on the front end and charge merchants on the back end because merchants receive a person, and the marginal cost of serving a single query is extremely low.

Agents, however, have real marginal costs on the front end. Reasoning, virtual machines, browser rendering, and failed retries all cost money. On the back end, what's delivered to merchants is an order without an attached user relationship.

Meta, though, wants to have its cake and eat it too. It is currently charging users a subscription fee on one side while planning to charge merchants a commission on the other.

Muse currently offers three pricing tiers: a free tier with roughly 100 million tokens per week, a Power tier at $20 per month, and a Maximum tier at $100 per month. Meta's Chief AI Officer, Alexandr Wang, stated that most users find the free tier sufficient, and the subscription tiers are designed to help cover the compute costs of heavy users.

Zuckerberg, meanwhile, has revealed that Meta ultimately intends to take a very small cut from transactions facilitated by Muse. This fee could be borne by merchants, with Muse aiming to "help you make money and help you save money."

Putting these three things together reveals a fundamental contradiction.

Helping users save money means squeezing merchant margins or shifting share from one merchant to another. The thinner the merchant's profit, the less willing they are to pay commissions.

Subscriptions are only meant to cover compute costs for heavy users, implying the paying user base is destined to be small.

Institutions have already done the math.

Oppenheimer, which rates Meta as neutral, has estimated that with roughly 1.91 billion Muse users and a paid conversion rate of about 6%, similar to ChatGPT, that yields approximately 115 million paying users. At $20 per user per month, that's about $27.5 billion a year, corresponding to roughly a 20% boost to Meta's earnings per share.

The analyst himself doubts whether Muse can reach this level, citing paid conversion, competition, and user trust. His own estimate for ChatGPT's paying users is only 33 to 88 million.

Morgan Stanley's accounting is even less favorable. If Muse reaches 100 million users by 2028, with each user making 5 queries per day, 10% having commercial value, and monetizing at $0.07 per instance, that's approximately $1.3 billion in annual revenue, contributing only 1% to Meta's earnings per share that year.

Meta's second-quarter advertising revenue was $59.36 billion, annualized at roughly $237 billion. One institution's projection amounts to about 12% of the existing base, while another is less than 1%.

Even within Meta's revenue landscape, personal agent subscription fees are nothing more than a supporting role.

The Art of Choosing Sides

Every business model attempting to build a personal agent must answer a fundamental contradiction: does it serve consumers or merchants?

Muse's demonstrations help users compare prices, negotiate bills, or cancel subscriptions. Its actions consistently side with the payer, facing the payee on the other side. Whether the payee is willing to open the door for it determines how far it can go.

Many platforms positioned to serve consumers have historically hit business model walls, with shopping guides being the most typical example.

Smzdm, a content-driven shopping guide platform that has operated for over a decade with modest prominence, is one of the purer buyer-side tools. It helps users compare prices, find deals, and make purchase decisions. Of its annual revenue of 1.288 billion yuan, information promotion accounts for about 40%, operational service fees about 33%, and performance marketing about 24%.

Information promotion includes e-commerce referral commissions and ad displays, performance marketing charges advertisers based on results, and operational service fees involve running store operations for brands. A platform that speaks for buyers still needs to survive on sellers' money. This business model contradiction caps its potential.

Big tech firms have also tried price comparison tools.

In late 2010, Alibaba launched Yitao, which aggregated prices from across the web. Before long, JD.com modified its crawler rules to block Yitao. JD was then boosting its gross margins ahead of its IPO, and Yitao's price comparisons directly pressured its pricing. Clearly, no platform would sit idly by while a price comparison tool helped users negotiate down prices on its own shelves.

Google had earlier launched a price comparison index, Froogle, in 2002, where any product could be listed and compared. After several name changes, it became Google Shopping in 2012, where merchants had to pay for visibility. The buyer's tool had turned back into a seller's tool.

The dilemma facing agents is a direct continuation of this lineage.

Standing with users to save them money makes it difficult for platforms that earn from traffic and ads to cooperate.

On September 20, Amazon's announcement that it had blocked Muse was no coincidence. Amazon cited unauthorized access, browsing without disclosing its AI identity, and the apparent capture and storage of user credentials. Amazon stated that third-party apps placing orders on behalf of customers with other merchants should operate transparently and respect the service provider's decision on whether to participate.

If an agent pivots to charging merchants for referrals, it ceases to be a buyer-side agent saving users money and becomes just another advertising channel.

Of course, there are those actively embracing Muse, like Shopify. This is because Shopify primarily profits from website-building tools and payments, and customer relationships belong to the merchants. Every transaction an agent brings is likely incremental for its small and medium-sized merchants.

Amazon is different. It has its own storefront and ad business. Users would naturally come to order directly. An agent stepping in between is siphoning off the exposure that would have occurred on its homepage and search results pages.

Meta currently wants both the apple and the pear.

Zuckerberg says Muse aims to "help users make money and save money," yet the revenue relies on merchant transaction commissions. He then also mentions that Muse can connect to Meta's advertising system, helping users run their businesses and place ads. This is precisely because nearly all of Meta's revenue comes from advertisers.

Saving users money and charging merchants are directly opposed in interest. This game will eventually force a choice of sides.

The Shelf Life of a Gateway

Something that can become a gateway needs at least two things: users can't live without it, and competitors can't easily copy it. In short, it requires sufficient differentiation.

Standard protocols like MCP are rapidly homogenizing agents' ability to connect to Gmail, calendars, maps, and music. Any agent can help users check emails, view schedules, or book restaurants.

If the personal agent path truly works, the most likely frontrunners should be Google, which holds Gmail, Chrome, and Android, or OpenAI or Doubao with the most users. It shouldn't be Meta, which only has social networks.

In fact, they've already done it.

OpenAI launched a ChatGPT agent in July 2025, allowing models to operate web pages on their own virtual computers.

Google unveiled Gemini Spark at its I/O conference on May 19 this year. It too runs in the cloud, acting as a round-the-clock personal agent with native access to Gmail and Calendar, launched more than three months earlier than Muse. A personal cloud virtual machine with a visible browser—this architecture is not unique to Meta.

Nat Friedman, head of product at Meta's Superintelligence Labs, has admitted that Muse draws heavy inspiration from the open-source agent OpenClaw. What they aim to build is a safe, user-friendly OpenClaw that can scale to billions of people.

What truly differentiates them are two things.

First, price. Gemini Spark is only available to subscribers of the $100-per-month AI Ultra plan. Muse offers 100 million free tokens weekly, plus a complimentary virtual machine.

In consumer scenarios, users are naturally price-sensitive. Muse's generous free tier is an advantage, but if other giants follow suit and compete, it will only raise the cost of sustaining this business.

Interestingly, Muse's most praised use case in the US is built on Google's assets. More reviewers believe its best use is connecting all their Gmail accounts. Even Meta's demos, like filling out school forms or tracking airline refunds, require first reading the user's email.

A product leading through free distribution builds its most valuable capability on a competitor's email service. For Google to catch up, all it needs to change is its price list.

Second, distribution. Muse works directly within WhatsApp, and Meta's family of apps has 3.6 billion daily active users.

Whether it's price or channel, none of these factors are about product strength. They rely on the profit "health bar" of other business lines. This is the same playbook as Qwen's 3 billion yuan Spring Festival giveaway, just swapping milk tea subsidies for compute power.

Moreover, Muse has won the starting race for a new product but hasn't disrupted the existing landscape. Sensor Tower data, cited by CNBC, shows that during the same 13-day period after Muse's launch, ChatGPT's downloads were around 3.1 million, still more than Muse. Claude and Grok had only about 400,000 and 200,000 downloads, respectively.

Many companies that previously experimented with personal agents have largely moved on.

OpenAI handed checkout back to merchants and shifted to product discovery. Doubao has pinned its monetization hopes on productivity scenarios like PPT creation and data analysis. Doubao Phone has retreated from "breaking down the door" to "knocking on the door." Each time they err, they move their revenue collection point to a more certain location.

There is an argument that personal agents can accumulate conversational memory and preferences, creating switching costs. However, through practical testing, we've found that neither personal data export nor MCP connectivity constitutes a switching cost for applications like Muse.

Another way to look at it: a personal agent as a 2C application has low defensibility. It's difficult to form network effects similar to social, e-commerce, or short-video platforms.

What locks users to the internet is often the legacy assets behind the agent: Meta's social graph, WeChat's relationship chain, Taobao's purchase history and payment accounts, Google's email. The agent is merely a feature parasitizing these legacy assets.

For Meta, Muse's emergence certainly has meaning. It can refine Meta's transaction ecosystem. The agent gives it a chance to intercept some consumer purchase intent from Google and Amazon. Likewise, it needs a story for the $130 to $145 billion in capital expenditure.

After Muse topped the charts, Wells Fargo analyst Ken Gawrelski said Meta "now has a story to tell."

But its second-quarter free cash flow was a mere $784 million, zero share buybacks in the first half of the year, and long-term debt increased by about $25 billion in six months.

A narrative is a narrative, and a gateway is a gateway. The capital markets price the narrative, but products must earn money through the gateway.

Resistance to Replication

After Muse went viral, domestic discussions quickly turned to who the Chinese version of Muse would be.

But before solving that problem, it's necessary to recognize several key differences in the internet environments of China and the US.

First, the infrastructure states differ fundamentally.

The US has neutral e-commerce infrastructure like Shopify and open payment interfaces like Stripe. Merchants willing to be called by agents can integrate easily.

In China, e-commerce and local services are run by a few giants independently. Taobao, JD, Pinduoduo, Meituan, and Douyin each have their own product catalogs, pricing systems, and payment tools.

For an agent to help users compare prices and place orders across platforms, the first hurdle is whether those platforms are willing to open their doors.

Doubao has already hit this wall twice.

In December 2025, a technical preview of the Doubao Phone Assistant launched with Nubia phones. It used system-level permissions and simulated clicks to operate other apps but was quickly restricted by apps like WeChat and Taobao. Users' WeChat accounts were even forcibly logged out.

On September 14 this year, the consumer version of the Doubao Phone Assistant launched, introducing a screen automation operation statement protocol called SAEP. The public comment period runs until October 15, during which it only operates on system apps, ByteDance-owned apps, and explicitly approved third-party apps. Practical tests show that in-app automation for WeChat, Taobao, Meituan, and Xiaohongshu remains unavailable.

Second, differences in price sensitivity must be considered.

Chinese internet users are accustomed to free products, and fewer are willing to pay monthly subscriptions for AI tools compared to Americans.

Qwen's basic services are free. Doubao launched a three-tier paid subscription starting at 68 yuan at the end of June, but daily conversations remain free. Qwen's 3 billion yuan Spring Festival subsidy campaign was also far more aggressive than Muse's free quota.

Third, user habits differ.

Americans conduct a large volume of formal correspondence via email: flight and hotel confirmations, bills, banking and insurance notices, school forms—all ultimately land in the inbox, forming a comprehensive ledger of one's digital life. When an agent gets Gmail authorization, it gains access to this ledger, making context aggregation especially efficient. Most of the praise for Muse's product centers on this aspect.

In China, daily communication happens on WeChat, work on DingTalk or Feishu. Flight changes, delivery, and bill reminders are scattered across SMS, app push notifications, and service accounts. Refunds or ticket changes are usually done with a few taps in an app or through customer service. Super apps have already compressed most life services into one or two interfaces. Users can often do things themselves in a minute or two, leaving agents little time to save.

The habit of paying monthly for software is also much weaker. Aside from video and music memberships, users paying monthly for tools remain a minority. For an agent to persuade Chinese users to subscribe by claiming "the money saved monthly exceeds the subscription fee," there simply aren't that many subscriptions to save on. This means the charging space for personal agents in China is narrower than in the US.

How Giants Choose to Follow

Nevertheless, personal agents as a conceptual possibility continue to be actively explored by major domestic tech firms.

Over the past six months, their approaches have generally fallen into several paradigms.

The first is building an agent within their own empire's walled garden.

On May 11, Alibaba announced full integration between Qwen and Taobao. Taobao packaged its search, order fulfillment, and after-sales capabilities as "Skills" for Qwen to invoke. In June, Qwen opened up to third-party agents and skills, with Luckin Coffee, KFC, Mixue Bingcheng, and China Eastern Airlines among the first brands to test it.

But the products and services Qwen can call are mostly still within Alibaba's ecosystem. Third-party brands join as their own agents, carrying their own repurchase recommendations and membership reminders.

For users, this feels more like consolidating the Alibaba ecosystem into one interface rather than creating a new gateway.

For Alibaba, the benefit is that users don't need to leave Qwen to open the Taobao app. The risk is that agents bypass Taobao's homepage recommendation feed and search ad slots. When a user directly says "help me buy a carton of milk," they no longer see the carefully arranged promoted products on the homepage.

In Alibaba's quarterly report through June, the "AI Lab and Applications" division where Qwen resides was disclosed separately for the first time, showing an adjusted EBITA loss of 13.861 billion yuan, compared to a loss of 3.224 billion yuan in the same period last year. Alibaba attributed this to increased AI capability investments and rising inference costs for the Qwen app. Free cash flow for the same quarter was approximately negative 44.7 billion yuan. Three days after the earnings release, Alibaba announced a share placement to raise 80 billion Hong Kong dollars, all earmarked for AI.

The second approach is building a connector and ecosystem, typified by WeChat. This might be the closest existing counterpart to the Muse narrative.

On June 8, WeChat published guidelines for developers to access its AI ecosystem. JD, Meituan, Ctrip, Didi, and Dewu subsequently announced integrations. On June 17, WeChat Pay launched an AI-exclusive card for agents. On June 20, Xiaowei began grayscale testing.

Xiaowei's technical route involves calling mini-programs through WeChat's open platform, without reading the screen or simulating clicks.

Xiaowei remains in grayscale testing. Martin Lau stated on the second-quarter earnings call that Xiaowei's prototype can technically handle complex agent processes, but it is currently set to require user involvement and multi-step confirmations as a safety measure.

One explanation is safety considerations. Another, more direct explanation is this: Tencent's second-quarter marketing services revenue was 43.565 billion yuan, up 22% year-on-year, with the company attributing growth to AI-driven ad recommendation models, intelligent delivery products, and closed-loop marketing within the WeChat ecosystem.

These revenues grow from the attention on Channels, Moments, and Official Accounts. If Xiaowei lets users skip browsing entirely and directly say "help me order takeout on Meituan," the value of a portion of that ad exposure diminishes. An analyst on the earnings call has already posed this question: if transaction paths shorten, would it shrink the high-margin ad inventory along with it? Lau's explanation was that these risks don't hold, and AI will make the WeChat ecosystem more valuable.

Xiaowei's restraint may precisely stem from WeChat's clear understanding of what drives its revenue.

The third approach is being called upon, with Meituan as a representative.

On June 1, Wang Xing announced on the first-quarter earnings call that Meituan's AI assistant, "Xiaomei," would integrate with Tencent's Yuanbao. Users could raise local life service requests within Yuanbao, and Xiaomei would handle the transactions. He also introduced the "To A" concept: beyond serving consumers and merchants, serving AI agents is becoming increasingly important.

Meituan has its own Xiaomei and Xiaotuan but hasn't built a cross-platform general-purpose agent. Instead, it packages capabilities like food delivery, hotel & travel, and in-store services, waiting to be called upon by various agents. Transactions, pricing, subsidies, and user data all remain in Meituan's hands. The stance is to open interfaces while keeping the cash register.

The fourth path is the most aggressive in opening permissions: Doubao.

As China's most widely covered personal agent, it began attaching Douyin Mall product cards to conversations in October. It had already started internal testing an in-app ordering loop even earlier, while also working on a general-purpose agent at the phone system level.

But Doubao faces a paradox similar to Qwen. Douyin E-commerce's core strength is interest-based recommendation, using algorithms to turn people who don't know what they want into buyers. Once users skip the information feed and directly tell Doubao "help me buy a pair of running shoes," the recommendation value of the feed evaporates.

There is also a noteworthy absentee: Pinduoduo.

Pinduoduo quietly launched a natural language "AI search" feature at the end of May this year. However, to date, it has not released a consumer-facing shopping agent, nor has it announced integration with other major tech firms' agent platforms.

Among domestic e-commerce platforms, Pinduoduo is one of the most reliant on recommendation feeds. An agent means users bring clear intent and skip the feed to reach products directly. This could impact Pinduoduo more than any other platform.

The companies that depend most on "browsing" to make money are the quietest when it comes to agents.

Two Kinds of Scores, Two Kinds of Rewards

For big tech firms building personal agents, technology and product experience are far from being the challenge. The most dangerous outcome is realizing they are cannibalizing their own revenue.

Taobao sells merchants display slots through recommendation feeds and search ads. Every extra second a user stays on the page is an extra impression opportunity. Douyin's interest e-commerce relies on short videos and live streaming to create impulse purchases. WeChat's ads grow on the attention of Channels and Moments.

But the better the personal agent works, the less valuable these platforms' most profitable ad impressions become.

Therefore, those in China who should be most cautious about the personal agent narrative are precisely those best positioned to build it.

They have the users, data, and technology, but they lack a rational reason to use an agent to undermine their most profitable business.

For Meta, Muse feels more like an offensive weapon. Meta has no e-commerce transaction ecosystem of its own. The agent intercepts intent from Google and Amazon, and the incremental gains outweigh the cannibalization.

Even so, the most valuable portion of Muse's demand still originates from Instagram's recommendation feed. It fills in the execution but does not replace the feed.

Most Chinese giants possess both advertising and transaction businesses. The incremental gains from an agent versus the ad revenue it cannibalizes require careful accounting.

Alibaba's "AI Lab and Applications" lost 13.8 billion yuan in a single quarter, and Tencent's second-quarter capital expenditure grew 176% year-on-year. Whether these investments can generate new value exceeding the revenue they cannibalize is a question no company has answered yet.

In 2023, Zuckerberg cut capital expenditure, laid off staff en masse, declared the "year of efficiency," and Meta rose nearly 200%.

In 2026, he pushed capital expenditure to $130-145 billion, saw second-quarter free cash flow fall below $800 million, zero buybacks in the first half, and roughly $25 billion in new debt over six months. The market rewarded him with another 11% because of Muse.

The market doesn't care whether a company spends or saves. It cares whether the story holds together.

The "year of efficiency" story was: money spent right, excess cut. The Muse story is: money has somewhere to go, and a consumer gateway seems to be arriving.

The two narratives point in opposite directions, but both satisfied the market's needs at specific moments.

A personal agent is a good feature. It can make existing apps more useful and generate some new demand. But in the short term, it is unlikely to become a new gateway or support a wholesale repricing.

A gateway must generate demand, hold the ability to charge, keep users within its walls, and at least appear neutral in the eyes of users. Today's personal agents can claim at most one or two out of these four qualities.

Every minute of attention it saves is an asset internet companies could have otherwise sold. Every user it keeps is a person merchants would have been willing to pay to reach.

Morgan Stanley's chief internet analyst, Brian Nowak, said at the end of an internal discussion that what he's watching most closely is "whether users are actually buying things on Muse."

He doesn't care about downloads or daily active users, only transactions. The value of a gateway is measured by how many people pay and whether they return. The number of people walking in says nothing.

The current data is far from sufficient to answer that question, and Meta has yet to disclose more details on Muse's DAU, retention, or paid conversion.

Even if those numbers look great, the constraints it faces won't disappear.

Merchants are unwilling to pay a premium for orders without visibility into the user. Platforms refuse to be circumvented. Most user demand doesn't originate within the agent. And the agent layer can be replicated by any competitor.

Over the past few years, the AI narrative has been dominant, and markets have consistently favored revaluing tech companies based on their spending intensity.

But in many gateway narratives that never materialize, the scarcer capability has never been the courage to spend. It is the restraint to understand which money should never be spent.