Doubao and WorkBuddy, the showdown over "working hours"
A few days ago, the latest internal speeches of Zhang Yiming and Ma Huateng were exposed simultaneously.
Zhang Yiming talked about technology. He cheered ByteDance's AI team up, stating that the team has the capability to compete with leading players such as Anthropic and OpenAI at the technological frontier, and reaffirmed that ByteDance should not rely on distillation, but focus on long-term strategies.
Ma Huateng talked about the team. His speech conveyed a stronger sense of crisis and carried a more obvious warning tone, calling for "major adjustments to both personnel and organizations", and proposed a new playbook of "50% land force + 50% naval force".
After the public calls, what specific steps should be taken next?
The most notable move at present is that both companies are redeploying their core teams to compete for the high-value battlefield of office Agent.
ByteDance's spearhead is Doubao Work, while Tencent's is WorkBuddy.
Both products were launched only a few months ago, but they have received high priority across the entire companies and obtained a large amount of resource allocation. Liang Rubo and Ma Huateng have respectively stood behind and spoken for the two products, making them veritable "CEO-level projects".
In early November, the two key players entered a new round of competition, with the keyword being "openness".
Doubao Work has fully integrated third-party models such as Kimi, GLM, and DeepSeek, which run side by side with its native Doubao model, allowing users to choose freely.
On November 16, Doubao Work also launched the "Workbench" feature, which initially covers 21 professional scenarios including finance, healthcare, and education. It pre-configures skills and tools required for different professional scenarios to lower the threshold for AI usage, and shares similarities with the "Projects" feature of WorkBuddy.
As for WorkBuddy, it integrates WeCom and Tencent's office ecosystem internally, and is compatible with third-party IM and enterprises' own business systems externally. It will further open up its product architecture to support access to enterprise self-developed Agents, third-party Agents and third-party communication systems.
As for models, it has long provided a rich selection of options, not limited to Tencent Hy.
The two products convey the same attitude:
As long as users are willing to use my office Agent, it is entirely up to users to decide whether to use my models, Agents or office software.
The "Hill 597.9" of office Agent has become the top battlefield between ByteDance and Tencent.
1
Office Agent is not the core battlefield of the global AI industry, as leading companies still place more emphasis on foundational models, especially Anthropic and OpenAI.
ByteDance and Tencent, however, need to take a different path and bet heavily on office Agent.
The direct reason is simple: they have realized that their self-developed models lag behind SOTA, and cannot catch up in the short term.
As Zhang Yiming said, ByteDance has the ability to compete with OpenAI and Anthropic, but the earliest time to catch up with the cutting-edge technology will be 2029 or 2030.
As for Tencent, its Hy model has made remarkable progress after Yao Shunyu took charge of the team, but it is still not globally leading.
Tencent stated in its announcement: "Although Hy4 is already one of the strongest open-source models at present, there is still a significant gap between it and top closed-source models, especially in internally built high-difficulty test sets for cutting-edge mathematics, long-horizon development, model iteration, visual question answering and other fields."
It is not hard to see that the focus battle of the two giants at this stage is not to overwhelm each other's models - there is no point in two underperforming models competing against each other; instead, it is the application battle that traditional internet companies are good at.
The most worthwhile AI application to develop right now is office Agent.
Among all types of Agents, office Agent is one of the few applications that can generate clear output and achieve positive returns, and its usage volume has been rising rapidly.
According to estimates from Analysys, in September this year, the visits to 20 mainstream domestic desktop office Agents exceeded 300 million, a 4-fold increase compared to June. This growth rate is unreachable for all internet products, including the wildly popular free short dramas that are trending now.
At the same time, office Agent charges by seat: one enterprise customer needs to purchase dozens or even hundreds of employee accounts, with a higher unit price. Its ARPU is an order of magnitude higher than C-end AI subscriptions, with much broader business prospects.
Apart from direct monetization, another monetization scenario for office Agent is: building brand reputation among programmers, professionals and enterprises, so as to drive the growth of AI cloud services.
Office scenarios have a large number of complex, long-horizon, multi-modal tasks, and the token consumption far exceeds that of other scenarios. Therefore, taking control of the AI office entry through office Agent is equivalent to directly grasping the traffic valve of office tokens, which in turn drives the revenue growth of AI cloud that provides MaaS services.
The two giants have already realized this point.
ByteDance previously transferred the marketing, sales and customer service teams of Feishu to Volcano Engine, and one of the goals is to strengthen the collaboration between office scenarios and AI cloud.
Tencent, on the other hand, premiered its latest Hy4 model on WorkBuddy, and the effect was remarkable - a few days later, the surging traffic forced WorkBuddy to expand its inference cluster by 10 times urgently.
It can be said that occupying office Agent means occupying the first foothold of the token economy.
In this competition, the two giants have also formed new cognition: models are not the real moat.
To retain users, office Agent cannot simply rely on model capabilities, because new models will keep iterating, and no one can guarantee that they can always provide the strongest model at the lowest price.
Its real barrier is the accumulated office context over time: internal enterprise workflows, exclusive knowledge, chat records, emails, documents, schedules, meetings and so on. The key reason why users are willing to keep using the product and pay for it is that the more they use it, the more inseparable they become, and the migration cost gradually rises.
One piece of supporting evidence is that in the first quarter of this year, the active user retention rate of Tencent's WorkBuddy and CodeBuddy exceeded 60%, and the paying user retention rate exceeded 80% - these users obviously did not stay for the performance of the Hy model.
In this context, if ByteDance and Tencent rigidly stick to their own models, they will not only fail to form competitive advantages, but also fall behind the more open and inclusive party.
WorkBuddy chose to be compatible with diverse models very early. Doubao Work opened third-party models for teams and enterprise customers at the end of October, and fully opened this feature in early November.
Neither of the two office Agents takes models as an access threshold. It would not be surprising if one day Doubao Work integrates Hy and WorkBuddy integrates the Doubao model.
This choice shares a lot in common with the wildly popular DeepSeek Harness.
Although the three are not the same type of product, they all put models in a "drop-down menu" that users can switch at any time. They move forward along different paths, but independently reach the same conclusion: models are being downgraded from products to components; only by seizing more "working time" can companies occupy the high ground in competition.
2
The office Agent war is a rare "ByteDance chasing Tencent" scenario in recent years.
ByteDance has long outperformed Tencent in AI applications. For products including Doubao, Jimeng, as well as AI education and AI Coding products, ByteDance has always maintained a firm leading position, while Tencent's similar products rarely appear at the top of industry rankings.
WorkBuddy is one of Tencent's few AI products that far outperforms ByteDance's counterparts.
It was launched in March this year, seizing a 5-month market gap. Two months later, Tencent claimed in its financial report that it is the most popular efficiency AI agent service in China.
At that time, ByteDance's layout for office Agent was not comprehensive enough, mainly relying on TRAE (focused on AI Coding) and Coze to support its presence, and was obviously at a disadvantage in the competition.
In July, a report from Analysys stated that Tencent's office Agent ecosystem centered on WorkBuddy recorded about 32 million desktop visits, while ByteDance's side only recorded 14 million.
In the following months, WorkBuddy integrated QClaw, and achieved full interoperability with WeCom, further enhancing its function matrix and social capabilities.
In comparison, ByteDance did not launch the professional version of Doubao until the end of June. At the end of July, Doubao absorbed the Feishu product team, and at the end of August, it integrated TRAE and Coze to officially launch Doubao Work.
However, the gap between the two is still widening. In September, the desktop visits of WorkBuddy rose to 85 million, while that of Doubao Work exceeded 30 million. Although it has more than doubled, it is further away from the former instead of getting closer.
ByteDance obviously cannot accept this situation. The management started to make public calls.
At the all-hands mid-year meeting in early August, Liang Rubo listed Doubao as a "core" business. At the Feishu Conference in mid-September, Liang Rubo stated again that ByteDance will invest more resources and efforts in the enterprise market.
Facing internal and external pressure, Doubao Work has to take drastic measures.
One of the drastic measures is to give users the right to choose models, and voluntarily position itself as an Agent workbench. But this is far from enough - WorkBuddy has already done that. It needs to come up with something different from its competitor.
For example, token discounts. In early November, WorkBuddy canceled the 50% points discount for DeepSeek V4.1 Flash as planned, and restored the original price.
A week later, for this popular model, Doubao launched a limited-time discount of 0.01x consumption, which is basically equivalent to free usage. Although this small move has limited impact, it reflects Doubao Work's determination to win.
However, relying only on model price cuts is obviously not enough to win the whole battle. If Doubao Work wants to catch up with WorkBuddy, ByteDance must ultimately make good use of the Doubao App card.
At present, Doubao's MAU has reached as high as 419 million, far ahead of other competitors, but it has not brought user inflow to Doubao Work.
According to estimates from QuestMobile, the desktop MAU of Doubao Work in September was only 2.049 million, equivalent to 0.5% of Doubao App's MAU. In the same period, the MAU of WorkBuddy across three terminals (desktop, web and mobile) exceeded 32 million.
Of course, strictly speaking, this comparison is not fair: Doubao Work only has a desktop version, and it was just launched for one month at that time, but the monthly active user gap of more than ten times is still too large.
As the "parent body" of ByteDance's AI application ecosystem, Doubao is the decisive factor for the success of Doubao Work.
On the desktop side, Doubao is already very similar to Doubao Work, and the UI and function matrices of the two are becoming more and more alike. The real big part is the App side: Doubao Work does not have an independent App yet. No matter how large the traffic and how many users Doubao App has, it cannot be directly converted into downloads of Doubao Work.
In the past three months, Doubao Work has mainly been catching up - improving its organization, capabilities, models and pricing, and its product form is getting closer and closer to WorkBuddy. But for the most critical lesson of user volume, Doubao Work must get help from Doubao App to make up for it.
3
For office Agent, the two giants each have their own advantages, and both have shown the attitude of investing a large amount of resources. This also means that some businesses that do not show promising prospects for the time being will find it difficult to obtain group-level investment.
Among them is Personal Agent.
For a period of time in the past, driven by Meta Muse, foreign giants have successively entered the Personal Agent track. There are also many movements in China: ByteDance, Tencent and Alibaba all have related plans.
Especially for ByteDance, it was reported not long ago that ByteDance's Personal Agent is expected to be officially released in the coming weeks. This planned product will benchmark Meta Muse, but it is still unclear whether ByteDance will follow Meta to launch supporting hardware.
According to people familiar with the matter, Zhang Yiming recently introduced this plan at an all-hands meeting for the AI team, and clarified the key priorities.
But up to now, neither ByteDance's version of Muse nor the Personal Agent version of Yuanbao has come true. The three giants once showed the posture of launching a big battle for Personal Agent, but now they are holding back.
This may indicate that the management of ByteDance and other giants have doubts about Personal Agent.
Personal Agent does not meet the actual needs of China's internet. The operations of Muse that amaze Americans, such as auto insurance price comparison, subscription cleaning, and refund application, have long been part of Chinese netizens' daily life, and the related services are more mature, evolving to advanced features such as auto-renewal reminders, one-click unsubscription, automatic price protection refund, and "refund only" services.
If encountering such problems, the first reaction of Chinese users is to open the relevant App to contact customer service, and they can even file complaints online if the problem cannot be solved. If users ask third-party AI to help solve these problems, it will be low in efficiency, slow in speed, and will incur extra token costs.
The golden decade of the mobile internet era has allowed users, enterprises and platforms to run in with each other, and jointly form the most developed and active online and offline commercial ecosystem in the world. Almost everything Personal Agent can do already has Apps and mini-programs occupying the corresponding ecological niche, with solid user mindshare and customer base.
In this ecosystem, ByteDance, Tencent and Alibaba are the ones that provide platforms and infrastructure, and take the largest share of the cake.
Another big problem of Personal Agent is: every minute the Agent saves for the user, the platform loses one minute of user usage time. As an efficiency tool, its underlying logic is to help users save time, eliminate the time and energy consumption of browsing websites and Apps, and improve decision-making accuracy and transaction efficiency.
It allows users to skip browsing Apps and various contents, and directly reach the target products and services, which is a fundamental blow to the current business model of internet giants. This is not the situation that giants want to see, and it is also the deep-seated reason why giants' attitude towards Personal Agent turned cold after being hot at first.
Therefore, the motivation and demand for ByteDance and its peers to promote Personal Agent are not solid enough from the very beginning.
What's more, the giants have no extra energy to spare at this stage, and must use their limited resources on the most critical areas.
In the AI map of the giants, apart from office Agent, foundational models, AI + hardware and other tracks all have their own strategic value, and their business models are very clear. If they abandon everything else and forcefully develop Personal Agent, they will get half the result with twice the effort, and even damage their business foundation.
Doubao Work and WorkBuddy are the hills that ByteDance and Tencent are more eager to capture.
But the problem now is that neither of the two giants has a unique competitive weapon. What WorkBuddy can do - such as integrating third-party models, Doubao Work can also do, the only differences are the progress, ecosystem and software.
This has led to a slightly awkward situation: this major battle for the AI era uses a very "retro" playbook. The two sides are fighting secretly, but they have to compete for users by the intensity of model discounts. The office Agent war has gone back to the old path of the internet