Doubao Work and WorkBuddy Battle for the Office Agent Arena
ByteDance's Doubao Work and Tencent's WorkBuddy are competing around office Agents, and both are opening up access to third-party models. In November, Doubao Work integrated models such as Kimi, GLM, and DeepSeek, and on November 16 it launched workstation features covering 21 professional scenarios including finance, healthcare, and education; WorkBuddy, meanwhile, integrates the WeCom ecosystem and supports third-party Agent access. Yiguan calculations show that in September, visits to 20 mainstream desktop office Agents in China exceeded 300 million, a 4x increase compared with June.
Why read this
AI score 53
From model choice, preset scenarios, and traffic data, we can observe the competitive paths of the two office Agents and see how major tech companies approach applications after reducing models to components.
Topics
Body · AI translation
By Letter List
A few days ago, the latest internal speeches by Zhang Yiming and Ma Huateng were exposed at the same time.
Zhang Yiming talked about technology. He rallied ByteDance's AI team, believing they have the ability to compete with top players such as Anthropic and OpenAI at the technological frontier, while reiterating that ByteDance should not rely on distillation and should instead focus on long-term strategy.
Ma Huateng talked about the team. His remarks revealed a stronger sense of crisis and carried a sharper tone of admonishment. He hoped that "both people and the organization need major adjustments" and also proposed a new approach of "50% army + 50% navy."
After these rallying cries, what specifically should be done?
At present, the most conspicuous move is that both companies are deploying their forces to compete for the highly valuable battlefield of office agents.
ByteDance's spearhead is Doubao Work, while Tencent's is WorkBuddy.
Both major products have been online for only a few months, yet they are highly valued by their entire companies and have received a large tilt of resources. Liang Rubo and Ma Huateng have respectively endorsed and spoken for them, making them veritable "CEO projects."
In early November, the two main players dueled again, with the keyword being "openness."
Doubao Work fully integrated third-party models such as Kimi, GLM, and DeepSeek, putting them on par with its own Doubao model, allowing users to choose freely.
On November 16, Doubao Work also launched a "Workbench" feature, initially covering 21 professional scenarios including finance, healthcare, and education. It presets the skills and tools required for different professional scenarios to lower the barrier to using AI, which is similar to WorkBuddy's "Projects" feature.
WorkBuddy, meanwhile, integrates internally with WeCom and Tencent's office ecosystem, and is externally compatible with third-party IM and enterprises' own business systems. It will further open up its product architecture to support access by enterprises' self-developed agents, third-party agents, and third-party communication systems.
As for models, the options it offers have always been abundant and are not limited to Tencent Hy.
Both convey the same attitude:
As long as you are willing to use my office agent, whether you use my model, my agent, or my office software is entirely left to the user to choose.
This "Shangganling" of office agents has become the number one battlefield for ByteDance and Tencent.
01
Office agents are not the main battlefield of the global AI industry; leading companies place more emphasis on foundation models, especially Anthropic and OpenAI.
ByteDance and Tencent, however, need to take a different path and bet on office agents.
The direct reason is simple: they have realized that their own models lag behind SOTA and cannot catch up in a short time.
As Zhang Yiming said, ByteDance has the ability to compete with O and A, but truly catching up to the technological frontier will take until 2029 or 2030 at the earliest.
As for Tencent, Hy has made significant progress under Yao Shunyu's leadership, but it cannot be called globally leading either.
It stated in an announcement: "Although Hy4 is already one of the strongest open-source models currently available, there is still a considerable gap between it and top closed-source models, especially on internally built high-difficulty test sets such as frontier mathematics, long-horizon development, model iteration, and visual question answering."
It is not hard to see that the focal battle for the two giants at this stage is not to overwhelm the other's model—"two weak chickens pecking at each other" is meaningless—but rather the battle of applications, at which established internet companies excel.
The AI application most worth pursuing right now is the office agent.
Among various agents, office agents are one of the few applications that can deliver clear output and make the numbers work, and usage has been rising rapidly.
According to Analysys calculations, in September this year, visits to 20 mainstream domestic desktop office agents exceeded 300 million, a fourfold increase compared with June. This is a growth rate that all internet products—including the currently booming free short dramas—find hard to match.
At the same time, office agents charge seat fees. An enterprise customer needs to purchase dozens or even hundreds of employee accounts, and the unit price is higher. Its ARPU is an order of magnitude higher than C-end AI subscriptions, and its commercial prospects are much broader.
In addition to direct charging, another monetization scenario for office agents is building brand reputation among programmers, professionals, and enterprises, thereby driving growth in AI cloud services.
Office scenarios contain a large number of complex, long-horizon, multimodal tasks, and token consumption far exceeds that of other scenarios. Therefore, controlling the AI office entrance through office agents is equivalent to directly controlling the traffic valve for office tokens, thereby driving revenue growth for AI clouds that provide MaaS services.
The two giants have already realized this.
Tencent gave the latest Hy4 model its debut in WorkBuddy, with astonishing results—a few days later, surging traffic prompted WorkBuddy's inference cluster to urgently expand tenfold.
It can be said that occupying office agents means occupying the first foothold of the token economy.
In this battle, both giants have also gained a new understanding: the model is not the real moat.
If an office agent wants to retain users, it cannot rely solely on model capabilities, because new models will always keep iterating, and no one can guarantee always being able to offer the strongest model at the lowest price.
Its real barrier is the accumulated office context: internal enterprise workflows, proprietary knowledge, chat records, emails, documents, schedules, meetings, and so on. The key reason users are willing to keep using and paying is that the more they use it, the less they can do without it, and switching costs gradually rise.
One piece of supporting evidence: in the first quarter of this year, Tencent WorkBuddy and CodeBuddy had active user retention rates above 60% and paid user retention rates above 80%—these users clearly did not stay because of the Hy model's performance.
Under these circumstances, if ByteDance and Tencent rigidly stick to their own models, they will not only fail to build a competitive advantage, but will also be left behind by the more open and inclusive side.
WorkBuddy chose inclusiveness early on. Doubao Work opened third-party models to team and enterprise customers at the end of October, and fully opened them in early November.
Neither of the two major office agents uses models as a threshold anymore. If one day Doubao Work integrates Hy and WorkBuddy integrates Doubao, it may not be surprising.
This choice has much in common with the hugely popular DeepSeek Harness.
Although the three are not the same type of product, they all put models into a "drop-down menu" that users can change at any time. They travel along different paths, yet independently reach the same conclusion: models are being downgraded from products to components; only by controlling more "work time" can one occupy the competitive high ground.
02
The battle over office agents is a relatively rare case in recent years of "ByteDance chasing Tencent."
WorkBuddy is one of the few Tencent AI products that leads ByteDance by a wide margin.
It launched in March this year, seizing a five-month market gap. Two months later, Tencent claimed in its financial report that it was China's most popular productivity AI agent service.
At that time, ByteDance's layout in office agents was not yet comprehensive, relying mainly on TRAE, which focuses on AI Coding, and Coze to hold the fort, and it was clearly at a disadvantage in the competition.
By July, a report by Analysys said that Tencent's office agents, with WorkBuddy at the core, had captured about 32 million desktop visits; ByteDance's side had only 14 million.
In the following months, WorkBuddy absorbed QClaw and fully connected with WeCom, and its feature matrix and social capabilities continued to strengthen.
By contrast, ByteDance did not launch Doubao Professional Edition 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, officially launching Doubao Work.
But the gap between the two continued to widen. In September, WorkBuddy's desktop visits rose to 85 million, while Doubao Work exceeded 30 million. Although the latter more than doubled, it actually fell further behind the former.
ByteDance clearly could not accept this situation. Management began to speak out publicly.
At the company-wide mid-year meeting in early August, Liang Rubo listed Doubao as a "backbone" business. At the Feishu conference in mid-September, Liang Rubo said again that ByteDance would invest more resources and strength in the enterprise market.
Facing internal and external pressure, Doubao Work had to take drastic measures.
One of those measures was to give model choice to users and voluntarily retreat to the position of an Agent workbench. But that was far from enough—WorkBuddy had already done so long ago. It also needed to offer something different from its rival.
For example, token discounts. In early November, WorkBuddy canceled the 50% credit 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 offer of 0.01x consumption, essentially equivalent to free use. Although this small move had limited impact, it also reflected Doubao Work's determination to win.
But relying only on model price cuts clearly cannot win the entire battle. If Doubao Work wants to catch up with WorkBuddy, ByteDance ultimately still needs to play the Doubao App card well.
At present, Doubao's MAU has reached as high as 419 million, far ahead of other rivals, but it has not injected users into Doubao Work.
According to QuestMobile estimates, in September Doubao Work's desktop MAU was only 2.049 million, equivalent to 0.5% of Doubao App. In the same period, WorkBuddy's MAU across three platforms (desktop, web, and mobile) exceeded 32 million.
Of course, strictly speaking, this comparison is not fair: Doubao Work only has a desktop version and had just been online for one month at that time, but a more than tenfold gap in monthly active users is still too large.
On desktop, Doubao is already almost on par with Doubao Work, and the two are increasingly similar in UI and feature matrix. The real bulk is still on the App side: Doubao Work still has no standalone App, and no matter how much traffic or how many users Doubao App has, it cannot directly translate into downloads for Doubao Work.
Over the past three months, Doubao Work has mainly been catching up—catching up on organization, capabilities, models, and pricing—and its product form is increasingly close to WorkBuddy. But on the most critical lesson, user volume, Doubao Work must have Doubao App help it catch up.
03
In building office agents, the two giants each have their own advantages and have both shown a posture of investing substantial resources. This also means that some businesses with no visible prospects for the time being will find it difficult to obtain group-level investment.
Among them is Personal Agent.
For some time, driven by Meta Muse, major foreign companies have entered the Personal Agent space one after another. There have also been many signs of activity in China, with ByteDance, Tencent, and Alibaba all having related plans.
ByteDance in particular: not long ago, reports said ByteDance's Personal Agent was expected to be officially released within the next few weeks. This planned product will benchmark against Meta Muse, but it is not yet clear whether ByteDance will follow Meta in launching supporting hardware.
But to this day, neither the ByteDance version of Muse nor the Personal Agent version of Yuanbao has become a reality. The three giants once struck the posture of a major battle over Personal Agent, yet now they are holding their fire.
This may indicate that the management of ByteDance and other major companies harbor doubts about Personal Agent.
Personal Agent does not fit the actual needs of China's internet. The operations in Muse that amazed Americans—such as comparing car insurance prices, cleaning up subscriptions, and requesting refunds—have long been everyday routines for Chinese netizens, and are done more maturely, having evolved into advanced practices such as automatic renewal reminders, one-click unsubscription, automatic price-protection refunds of differences, and "refund only."
If Chinese users encounter such problems, their first reaction is to open the relevant App and look for customer service; if that really doesn't work, they can also file complaints online. If a third-party AI were asked to help solve them, it would be inefficient and slow, and would also incur extra token costs.
A decade-long golden age of mobile internet has allowed users, enterprises, and platforms to adapt to one another and jointly form the world's most developed and active online-and-offline commercial ecosystem. Almost everything Personal Agent can do already has an App or mini-program occupying that ecological niche, with entrenched mindshare and customer bases.
In this ecosystem, the ones providing platforms and infrastructure and taking the largest slice of the pie are ByteDance, Tencent, and Alibaba.
Another major problem with Personal Agent is this: every minute an Agent saves the user is one minute less of user time for the platform. As an efficiency tool, its underlying logic is to help users save time, sparing them the time and energy spent scrolling websites and Apps, and improving decision precision and transaction efficiency.
It lets users skip past browsing Apps and various content and go straight to the target goods and services—a fundamental undermining of the major companies' current business models. This is not a situation the major companies want to see, and it is also the deep-seated reason why their attitude toward Personal Agent turned from hot to cold.
Therefore, the motives and needs behind ByteDance and the others vigorously promoting Personal Agent were not solid to begin with.
Moreover, at this stage the major companies have no time to attend to other matters and must apply their limited resources where they count most.
On the major companies' AI maps, apart from office Agents, large models, AI plus hardware, and others each have their own strategic value, and their business models are also very clear. If they abandoned the rest and mounted a strong attack on Personal Agent, they would only get half the results with twice the effort, and might even damage their commercial foundations.
Doubao Work and WorkBuddy are the hilltops that ByteDance and Tencent are more eager to conquer.
But the problem now is that neither major company can field a uniquely competitive weapon. What WorkBuddy can do—such as connecting to third-party models—Doubao Work can also do; the only differences are in progress, ecosystem, and software.
This has created a somewhat awkward situation: this major battle for the AI era is being fought in a very "retro" way. The two sides fight openly and maneuver in secret, yet they need to compete for users through the size of model discounts. The office Agent war has looped back to the old internet-era path of burning money to win users.
To break the deadlock of a zero-sum game, the major companies in the race may still have to return to foundation models—at this stage, everyone's foundation model matrices are about the same; only by producing an exclusive SOTA foundation model and putting it into office Agents at an affordable price will the real watershed moment for this battle arrive.