ByteDance's Doubao cannot develop Muse

The 30-day public notice period has ended. WeChat, Alipay, Taobao, Meituan, Pinduoduo, DiDi, and Xiaohongshu have all refused to grant access. What China's AI Agent has hit is a wall not built by technology.

When you wake up in the morning, you want your phone to do a few things for you: compare prices and place an order on Taobao, order lunch on Meituan, and book a ride on DiDi. Previously, you had to tap through all these steps manually. Now AI assistants claim they can do all these for you.

The Doubao mobile assistant from ByteDance once made exactly this promise. It aimed to become China's Meta Muse, China's Grok Bot — an intelligent agent that lives in your phone and can operate other apps on your behalf.

To achieve this goal, Doubao took a fairly rule-abiding step: it handed over the control switch to those apps themselves. On September 14, the Doubao mobile assistant launched an agent authorization public notification mechanism (internal code name SAEP), giving all cooperating apps a 30-day window to independently decide whether to allow Doubao to act as an agent to perform automatic operations on behalf of users. The window closed on October 14.

Looking at that list again today (October 15), a whole column of names is grayed out: WeChat, Alipay, Taobao, Meituan, Pinduoduo, DiDi, Xiaohongshu. None of them were left out.

This is not a small-scale test failure. It is a collective turnaround of almost all top-tier applications in China.

It is not that it cannot be built, but that it is not allowed to be built

One thing needs to be clarified first: Doubao's inability to become Muse has almost nothing to do with technical capabilities.

The work of tapping through apps, filling out forms, placing orders, and making payments on behalf of users has long been verified in engineering. Similar products overseas are already in operation — Meta's Muse embeds the assistant into social media and hardware, and Elon Musk's Grok is also being integrated into operating systems and hardware entry points. A more typical reference is OpenAI's Operator and Anthropic's Computer Use, which directly operate third-party websites for users in the browser or desktop environment, handling tasks such as shopping, form filling, and ticket booking with ease.

Doubao is not incapable of doing this. Visual understanding, interface parsing, and automated clicking are all standard features of current large model agents. What it lacks is not the "hand" to operate, but the willingness of others to open the "door".

This is exactly what the SAEP public notification mechanism was designed to solve: instead of forcing entry, it is better to issue an open invitation and let each app make its own choice. If they choose "Allow", Doubao can enter as an agent; if they choose "Reject", the door remains locked. In the 30-day window, no one chose "Allow".

Are American apps really more "easy to talk to"

When you shift your focus overseas, the situation is indeed quite different.

Operator can access a large number of websites to place orders in the US; Muse, backed by Meta's huge ecosystem, is naturally embedded in the entry points that users open every day; Grok, leveraging the partnership between xAI and hardware companies, is also seizing the position of "system-level assistant". On the surface, it seems that American app companies are more willing to issue access permissions to third-party agents.

But this comparison has its limitations. The US does not have completely open access — a large number of websites block automatic agents through robots.txt, anti-scraping policies, and login risk control mechanisms. Operator itself also runs in an isolated environment, and key operations require user confirmation. Websites in sensitive industries such as banking, aviation, and healthcare even block automatic agents as abnormal traffic. What Operator can really run smoothly are those standardized, low-risk e-commerce and form sites, far from the state of "being able to enter any store at will". What really allows American agents to "survive better" than their Chinese counterparts is not that large companies are more generous, but three things that are missing in China:

The first is the open tradition of the Web. Many services in the US exist in the form of websites, with open HTTP at the underlying layer and standard login and payment processes, so agents can access them by following the protocols. Almost all core services in China are enclosed in super apps, web pages are only supporting roles, and all real transactions, social interactions, and payments are fully closed-loop within apps.

The second is the external pressure of anti-monopoly regulations. The EU's Digital Markets Act (DMA) directly forces "gatekeepers" such as Apple and Meta to open up interoperability; although the US does not have such strict regulations, regulators' tolerance for platform blockades is declining. China does not have this kind of coercive force, and platforms can close their doors without facing any penalties.

The third is the clear boundary of responsibility and compliance. American agents operate under the framework of "user authorization + isolated environment + step-by-step confirmation", and the responsibility chain in case of accidents is relatively clear. For third-party agent operations in China, if an order is placed incorrectly, payment is made to the wrong recipient, or information is leaked, there is no existing clear rule on whether the responsibility falls on the platform or the agent.

Three walls standing in front of Doubao

Why can't China build a Muse? When you break it down, there are three very tangible walls.

The first wall is the power of entry points. China's internet has developed into a landscape of "super apps" — each of them is a traffic entry point and a trading platform. If Doubao can place an order on Taobao for you, then Doubao's chat interface will become the new entry point, and Taobao will be reduced to a back-end executor. No platform is willing to demote itself from the "front desk" to a "pipeline". This is not just an awkward situation for Doubao, but a zero-sum game that all players who want to build general-purpose agents have to face: whoever becomes the entry point wins; whoever becomes the execution layer will be sidelined.

The second wall is the lifeline of data. Every price comparison, every stay, and every order a user makes in your app is fuel for advertising, repurchasing, and model training. Allowing ByteDance's agent to operate on behalf of users is equivalent to feeding the most valuable user behavior data to the strongest local traffic competitor — Douyin E-commerce is already encroaching on Taobao and Meituan's market share. Rejecting Doubao is a business instinct, not a random emotional decision.

The third wall is regulation and trust. Proxy operations touch the three most sensitive areas: payment, accounts, and privacy. China has extremely strict regulations on payment security and personal information protection, and there is currently no clear compliance framework for third-party agents to operate on behalf of users. Not opening up access means at least no accidents; if access is granted, who will take responsibility if something goes wrong? This wall is not built by a single company, but by the entire compliance environment.

An app that acts simultaneously as an entry point, a data pool, and a strictly regulated payment node — when these three identities overlap, any of them can block third-party agents from entering, and the combined force of the three will lock the door from the inside.

It is not just Doubao, but the entire structure

If you take a broader view, you will find that this is not at all a setback for Doubao as a single product.

Baidu, Alibaba, and Tencent all have their own ambitions for building agents, but they are all trapped in their respective walled gardens. Alibaba can make its own apps interoperate smoothly — Taobao, Alipay, and Quark work well together, but WeChat will never connect to them. The same goes for Tencent: its own ecosystem works well, but it cannot enter other people's doors. China's internet has no history of "interoperability". The Web is inherently open, but China's digital life was later built as separate castles, and no gates were reserved when they were constructed.

When WeChat launched mini-programs back then, the original intention was to let services run lightly within WeChat, but it has always been closed to the outside world — can you open Taobao's mini-program in WeChat? No. The battle for entry points has been a zero-sum game from day one, not something that just emerged today.

What is even more interesting is another list: a certain mobile operating system also has an "intelligent agent open platform" that lists the access status of a series of apps, and the result is also full of rejections. Note that this is a mobile phone manufacturer, not ByteDance. Mobile phone manufacturers and apps have a "symbiotic" relationship: apps cannot function without mobile phones, and mobile phones cannot function without apps. Theoretically, a system-level agent from a mobile phone manufacturer has more leverage to negotiate for openness than a third-party app. Even mobile phone manufacturers have not reached an agreement, which shows that even the "system-level" path is currently blocked.

What does this mean for you

For ordinary users, this is a somewhat disappointing thing: the AI in your phone that claims to "help you do everything" will most likely get stuck at the prompt "the other party does not accept the agent" when it is time to place an order or make a payment for you. When you tell it "help me order takeout", it will reply "I have opened Meituan for you" — and then you still have to tap through all the steps yourself.

Is Muse still possible in China

Yes, but it is most likely not the kind of "general-purpose third-party agent" that Doubao envisioned.

One possibility is that regulators take the lead in promoting interoperability standards, like the EU's DMA, using rules to force platforms to open their doors. But China does not have such pressure at the moment, and this scenario cannot be seen in the short term.

Another possibility is that agents retreat to run in closed loops within a single giant's ecosystem — Alibaba's own ecosystem runs on its own, Tencent's own ecosystem runs on its own, and ByteDance's own ecosystem runs on its own. Cross-ecosystem general-purpose agents are a dead end under the current power structure. The "AI operating on my behalf" experience that users can actually enjoy will be cut into isolated islands, scattered in the built-in assistants of several super apps.

There is also a possibility that the mobile phone system layer will truly step up. If vivo, OPPO, Honor, and Xiaomi make their system-level intelligent agents the default entry point, apps will have to connect to them sooner or later to stay on the phones. But judging from the current rejection list, this step has just started, and it is far from being fully connected.

Final Notes

Doubao's inability to build Muse is not because the model is not powerful enough, but because China's internet has developed into a number of disconnected castles. The owner of each castle does not want to issue an access card to the butler of the neighboring castle. The 30-day public notice has simply made this fact completely clear to everyone.

This article is from the WeChat official account "Tech Business", author: Tech Business, published with authorization by 36Kr.