ByteDance Agent is waiting for the favorable opportunity

In just two months, all kinds of players have intensively shown their hands at the table of personal Agents.

Meta's Muse launched then faded, with its average daily net new users plummeting from 198,000 at the peak in September to 74,000, a drop of 62.4%; Alibaba's "Qianwen Go!" completed the mental registration of "handling affairs on behalf of users" by successfully settling a flight delay claim of only 70 yuan; Tencent's WeChat Mini Agent is still expanding its gray-scale testing coverage.

The three players have different styles and different progress, but at least they have all sat at the table.

While Xiaodou (Spell), the personal assistant that launched the internal test as early as April, is still "in internal test" so far, with "no public launch date announced", and there is not even a single observable case available.

ByteDance, which has long held a full hand of cards, has never shown its real strength, which is inconsistent with the public's perception of the company.

ByteDance used to be the most product-savvy company in China: in the era of graphic information feeds, Toutiao redefined content distribution with recommendation engines, and TikTok copied the same product mechanism to the whole world; in the short video track, Kuaishou took the lead, while Douyin came from behind and overtook it relying on sliding experience and music templates; even a tool product like CapCut has grown into the default option for short video editing.

This is an App factory that is used to achieving late-mover advantage and impacting the market with products and algorithms. However, in the Agent battle, ByteDance's efforts on both fronts have not achieved ideal results. Its personal assistant was preempted by "Qianwen Go!" to land in the market, and the office market was first occupied by WorkBuddy to grab a wave of traffic.

The reason may not lie in the product itself. ByteDance's core capabilities in the past focused on organizing products, algorithms and traffic into an efficient machine to complete cold start, user retention and scaling in the shortest possible time. But the competition of Agents lies in the upper limit of models, ecological authorization, and self-reform of business models, all of which put higher requirements on ByteDance's capability boundary.

I. Personal Agents Waiting for the Wind

Staying put means more considerations.

Personal Agent is exactly the track within ByteDance's range and with the best endowment: high frequency, no restriction on user groups, and to a certain extent, it can continue the network effect of spreading computing costs through scale. Morgan Stanley estimates that Muse's monthly cost per user is between 3 and 130 US dollars. Born at Meta, Muse's early outbreak exactly relied on the digital content ecosystem.

The user pool of Douyin and Doubao has endowments of the same level. For ByteDance, this is a battlefield where it can relatively reproduce the expansion ideas from the mobile Internet era. There is great opportunity in fully activating hundreds of millions of C-end users in the AI era.

Looking back, ByteDance has been "blocked" twice on the personal Agent battlefield.

The first time was at the end of last year. The first-generation Nubia Doubao mobile phone chose to break through the application blockade of competitors, trying to implement the concept of "making decisions on behalf of users": taking shopping as an example, it can automatically compare prices, search and place orders across Apps, and users only need to complete the final payment step. This mobile phone reached its peak right after release, and its price was even increased by 5000 yuan in the second-hand market for a period of time.

This "accessibility interface" route based on GUI Agent essentially reads the screen with system-level permissions and simulates manual clicks. It is equivalent to taking the user's eyes and fingers to forcibly open the access control of other companies' Apps.

No one will allow others to sleep soundly beside their bed. Less than a week after the carnival started, from the evening of December 2 last year, users who used AI to operate WeChat would trigger "abnormal environment" and be forced to log off; on December 5, Doubao shut down financial-related operations; on December 10, some Alibaba applications lifted the login restriction, on the condition that Doubao simultaneously closed the operation permission for related Apps.

With the experience of "hitting a wall", ByteDance chose to adopt a conciliatory approach before taking tough measures in the second battle. On September 14 this year, ByteDance launched the Screen Automated Operation Protocol (SAEP) along with the second-generation Doubao mobile phone (Nubia NaviX Ultra), leaving a 30-day window for all Apps to choose whether to allow Doubao to perform proxy operations, with fine-grained permissions and high-risk actions restricted by default.

When the public notice period expired on October 14, the result was that almost all of the 153 major Apps refused — WeChat, Alipay, Taobao, Meituan, Pinduoduo, Didi, and Xiaohongshu were not on the allowed list, even DeepSeek and Kimi also chose "no". The next day, ByteDance had to reverse the rule to "open by default, unless explicitly rejected".

So far, the two paths left for the fledgling personal Agent product "Xiaodou" are not easy.

From an external perspective, with core interests at stake, cross-application general assistants are blocked by external walls. Allowing ByteDance's Agent to enter their own ecosystem is equivalent to giving away the most core user behavior data and transaction entrances to competitors, and the vast majority of platforms will not allow this.

In this case, the development space left for Xiaodou may be to retreat to the Douyin ecosystem and become a "ByteDance system life assistant". Although the shelves are owned by ByteDance itself, if the Agent bypasses the recommendation feed to complete transactions directly, it will shake ByteDance's core foundation of advertising monetization.

To make the above conflicts acceptable, on August 10, the independent commission rate for travel and hotel transactions completed through the Doubao entrance (about 12%) is higher than the 8% on the main Douyin app, which can be regarded as a preventive measure for both internal and external stakeholders of the ecosystem.

This dilemma in business choice seems to make ByteDance, which has always been aggressive externally, hesitant for a while.

Qianwen Go! dares to make the first move, not because Alibaba has no advertising burden. Taobao and Tmall also rely on transaction commissions and merchant ads for revenue, but the consumption behaviors initiated by the Agent on behalf of users all occur within the ecosystem. The more successful the Agent is, the higher the efficiency of converting transactions into commissions, and the less erosion to ads; while ByteDance's closed loop is an "attention closed loop". Every minute the Agent saves for users, Douyin loses one minute of ad inventory.

When Xiaodou, the main Doubao app and Doubao Work are finally all under the leadership of Zhao Qi, and Xiaodou, which was exposed to be in development as early as April, has not yet been launched, the reason is very likely to stem from ByteDance's hesitation about its own business model.

II. The Tools Are Ready, But the Task Is Not Completed

At present, the situation of ByteDance's office Agent front seems more moderate.

When "Doubao Work" was released on August 25, WorkBuddy, which launched in March, had been leading the track for nearly half a year. Its PC monthly visits in June reached 20.97 million, ranking first in the industry. As the product was successively connected to Tencent Meeting, Tencent Docs, WeChat, WeCom and other products from the mobile Internet era, the context that the Agent can access and the actual use effect have been improved, and user awareness has also been gradually penetrated.

In the past, ByteDance was good at achieving late-mover advantage, and this time, ByteDance tried to speed up its catch-up with a lightning integration process that only took 25 days.

On July 30, the Feishu product team was merged into Doubao, Xie Xin's reporting line was transferred to Zhao Qi, and the marketing and sales team was merged into Volcano Engine; on August 24, the TRAE and Coze teams were merged into the Doubao "big family"; on August 25, Doubao Work was officially released and connected to the Feishu ecosystem.

It was not until September 15 that Feishu 8.0 was officially unveiled after the integration. Users can directly enter Doubao Work from the Feishu sidebar navigation bar, and it supports accessing Feishu functional products such as group chats and documents. At the same time, basic office services and collaboration functions are all integrated into Doubao Work. This means that the two sides have gone further in two-way integration, and completed full connectivity at the product level.

This extremely efficient integration has an unavoidable background of the times. At the product level, no AI-native application that can completely replace the tools of the previous era has been born yet; at the organizational level, the integration also solves the long-standing internal consumption problem — previously, Feishu, Doubao, TRAE and Coze fought their own battles, with scattered entrances and isolated data, even the same enterprise customer needed to connect to multiple different ByteDance solutions.

But solving internal consumption is a problem of "sharpening the tools", which does not mean "doing the thing well". At the underlying ecological mechanism level, ByteDance is encountering a deeper level of transformation and choice.

The so-called office Agent is the landing form of the Token economy on the enterprise side. Through the productized encapsulation of underlying Tokens, the pricing of Tokens is realized, which in turn drives consumption. As long as Token consumption occurs, service revenue will be generated.

Under the premise that "the same amount of user demand generates the same amount of Token consumption", opening up multiple models and sticking to self-developed models seem to make no essential difference in the scale of driven revenue on the surface, but they point to completely different paths in terms of technology evolution path and business competitiveness.

The advantage of an open ecosystem lies in the rapid iteration of engineering capabilities and traffic siphoning. Taking Tencent's WorkBuddy as an example, it was successively among the first batch to access the model Space Bunny (anonymous) and Step 5 Preview released by StepFun in the past month — without complicated API access, users can switch and set models with high freedom, which can be regarded as empowerment for users in the AI era.

Under this path, the open platform abandons the binding to a single base model, and puts all its focus on engineering scheduling, connector ecosystem and workflow encapsulation. The more open the platform is and the freer the interface is, the more Token consumption and traffic will naturally flow to it. Office Agents use this scarcity to gather the most active users.

Doubao Work, which chooses not to open up, competes on data closed loop and base model iteration.

Doubao Work, which adheres to the "closed suite" route and only carries the self-developed Seed model and Feishu context, pays the price of sacrificing strong attractiveness to users with diversified demands in the short term; the advantage is that all Token consumption and interaction scenarios are locked on its own Seed model, realizing the full amount of real data backflow, and continuously feeding and assisting the iteration of its own base model through evaluation feedback.

From this perspective, in order to support Zhang Yiming's strategic will of "doing its utmost to improve model capabilities", Doubao Work lets the office Agent undertake the function of a training ground to "supply materials and feed" the Seed model.

On the other hand, when enterprises procure office Agents, they are essentially buying a set of Tokenized services for organizational data, behind which the enterprise's organizational structure, permission system, documents, approval and collaboration relationships are all migrated as a whole.

The market once regarded "345 million monthly active users of Doubao" as ByteDance's key chip. However, the huge monthly active users of Doubao belong to C-end individual users, who can chat with AI, but are far from organizational payment.

Look at another set of data. DingTalk has 171 million MAUs, which can be directly supported by Qianwen Office, while Feishu has about 30.19 million MAUs, only about one sixth of DingTalk; WeCom relies on the WeChat relationship chain, so WorkBuddy does not need to rely much on the stock of its own collaboration platform, but penetrates along the WeChat-WeCom relationship chain.

Of course, ByteDance is not without opportunities. The July data of the AI product list shows that TRAE Work has 2.66 million monthly visits, and Coze has 2.17 million, with a month-on-month increase of 62%. But what is a sweet treat for developers may not be suitable for ordinary office workers. A person close to the TRAE Work team said, "It just copied the previous IDE system, which made the product hardly make any progress in the first half of the year."

The convergence of AI technology is forcing the evolution of tools, and unified and seamless interaction will bring more predictable opportunities.

In the AI era, "product is king" may no longer mean stuffing complex buttons into the interface, but using the most extreme underlying model and engineering capabilities, leaving the complexity to itself, and leaving the extremely simple "one-click delivery" to users.

III. The Two Ends of Technology Priority

Combining ByteDance's personal and office two fronts, we can clearly see that this company is shifting from the product-first approach familiar to people in the past to a longer and more expensive technology-first approach, which is also the key reason for its slowed rhythm.

In the past, Internet products often relied more on interactive interfaces with good experience, stronger recommendation algorithms and higher growth efficiency. But for Agents to truly complete tasks on behalf of users, the core competitiveness becomes the reliability of the model. Users can ask it to look up information, but may not dare to let the assistant pay, approve or process contracts, because that means handing over part of the decision-making power to AI.

To this end, Zhang Yiming set the goal of doing his utmost to improve model capabilities.

The adjustment of the organizational structure gives a clear signal: in February 2025, Wu Yonghui joined ByteDance from Google DeepMind; in October, Zhu Wenjia changed his reporting line to Wu, and Wu became the top leader in Seed research. Correspondingly, Qiao Mu left the company, Yang Jianchao took a break and was replaced by Zhou Chang who came from Tongyi Qianwen, and Li Hang transferred to the position of consultant.

Research-oriented leaders take the top positions, veteran product personnel step back, the data and reinforcement learning departments are re-strengthened, and new competition indicators have also changed.

Of course, basic model development is close to a long-term investment. Data, training, reinforcement learning and post-training all require time, computing power and talents, and it is difficult to verify returns through rapid launch and rapid trial and error like Internet products. Technology priority can raise the ceiling, but the process is destined to be long and extremely resource-consuming.

After resources are tilted towards technology, the convergence at the product level has not been completed synchronously. The entrance of Doubao Work is still unclear so far, and users can access it from the desktop client, Feishu sidebar, and the work mode in the Doubao App.

In addition to the existing office users, ByteDance seems to want to convert the individual users accumulated by Doubao into office users, but the existence of multiple entrances at the same time will not only disperse R&D, operation and enterprise service resources, but also blur user perception. The problem is: is this a personal assistant, an office suite, or an AI enhancement component for Feishu?

The entrance of office Agent is not just a button, but determines the direction where the organizational structure, permission system, documents and collaboration relationships finally settle.

In contrast, Qianwen clearly separates Qianwen Office from the original Qianwen entrance, targeting B-end users. According to the data disclosed by Alibaba in September 2026, when Qianwen Office had more than 30 million users in the first month of launch, enterprise users accounted for more than half (over 50%).

From Xiaodou for C-end users to Doubao for B-end users, ByteDance's ups and downs on the Agent battlefield are essentially an inevitable collision between the traditional business model and the technical laws of the new era.

In the traffic era, ByteDance was an invincible "efficiency machine", distributing attention through algorithms and realizing value through advertising; in the Agent era, the core of competition among companies has only become the certainty of the model, the openness of the ecosystem and the deep trust of organizational data.