The commercial conflict over office AI agents is unfolding in a remarkably straightforward manner.
Recently, during a visit to a residential elevator in Beijing's Chaoyang District, we discovered that ByteDance's TRAE Work had begun advertising. The ad reads, "For complex tasks, use TRAE Work," accompanied by a QR code offering new users 4,500 points upon scanning.
Not long ago, another screen in the same elevator was displaying ads for Tencent's WorkBuddy. This moment marks the latest escalation in the heated competition for office AI agents.
Behind the Scenes of the Elevator Clash
Over the past few months, major tech companies have largely completed their initial deployment of desktop office agents. Alibaba launched Qianwen Office, Tencent is marketing WorkBuddy, Baidu introduced DuMate, and ByteDance is pushing both Doubao Office, which is deeply integrated with Feishu, and TRAE Work, which is expanding into general-purpose office scenarios.
The tasks targeted by these products are becoming increasingly similar, but their underlying model strategies differ significantly. Qianwen Office anchors its capabilities on its own proprietary model system, while Doubao Office is backed by ByteDance's own models and the Feishu workflow. In contrast, TRAE Work and WorkBuddy emphasize a multi-model aggregation approach, aiming to combine the capabilities of different models into a single desktop workspace.
The proprietary model route requires continuous investment in training to enhance capabilities; the stronger the model, the higher the ceiling for tasks an agent can handle. The aggregation route, however, faces a different challenge: when you have multiple models at your disposal, which one should be used for which task?
There are currently no standard answers to these questions.
Elevator Ad Wars
In recent months, the competition among tech giants over office agents has been more visible in product launches and organizational restructuring. Among these, ByteDance's late July adjustments to its Doubao, Feishu, and Volcano Engine units drew particular attention.
Previously, Feishu operated as an independent BU under ByteDance, with its head, Xie Xin, reporting directly to CEO Liang Rubo. After the restructuring, the Feishu product team was integrated with the Doubao product team, and Xie Xin began reporting to Doubao's head, Zhao Qi. The Feishu GTM team was also further integrated with related Volcano Engine teams.
This restructuring was notable partly because Feishu's own business performance was not weak. Against this backdrop, the fact that Feishu's product and sales systems were still being reorganized and reassembled made ByteDance's intent to reallocate resources around AI-powered office solutions even more apparent.
Following the restructuring, the Doubao Enterprise Edition, which was developed with significant involvement from the Feishu team, began internal testing with some Feishu clients. Feishu's existing office capabilities, including documents, spreadsheets, meetings, and group chats, are now serving as the foundation for Doubao to penetrate further into enterprise workflows.
For a time, Doubao Office became the primary focus of external observers' attention on ByteDance's next moves in AI office products. However, it was TRAE Work that actually started competing with WorkBuddy first on the ground.
In the same elevator in a Chaoyang District residential complex, TRAE Work ads are now visible. Just a short time ago, the other screen in the same elevator was displaying ads for Tencent's WorkBuddy. Now, the WorkBuddy ads have been removed, and TRAE Work has appeared on the adjacent screen.
This encounter in the elevator has also brought ByteDance's alternative office agent strategy back into the spotlight. TRAE initially started as an AI coding tool. In June, TRAE SOLO was officially upgraded to TRAE Work, expanding its product scope from code generation to include research, writing, data analysis, plan development, and cross-team collaboration for general office scenarios.
The transition from SOLO to Work indicates that TRAE aims to cover more than just coding tasks. This product positioning is now being pushed towards a broader office user base. TRAE's expansion into general office use is not starting from scratch. According to data from Analysys, in June, the monthly desktop web visits for WorkBuddy reached 20.97 million, while TRAE's IDE domestic version had 12.79 million visits, ranking second. After its upgrade in June, TRAE Work achieved account interoperability and project resource sharing with TRAE IDE.
While Doubao Office garnered most of the attention from ByteDance's recent organizational changes, it is TRAE Work that is physically competing for users against WorkBuddy. TRAE Work is facing increasing competitive pressure externally, having to vie for users with WorkBuddy, and internally, it must also address how to coexist with Doubao Office.
The Subtle Boundaries in the Model Pool
Looking deeper into the model layer, the competition among office agents reveals another point of divergence. One path is to bet on proprietary models. Qianwen Office currently uses its own Qianwen model as a capability anchor. Alibaba defines it as being "based on the Qianwen series of large models," offering model selectors with different tiers like Basic, Advanced, Economy, and Qwen3.8-Max. Doubao Office's approach is also based on ByteDance's proprietary Seed models.
For such products, continuous training and iteration of the model's capabilities still directly determine the upper limit of the agent's performance. TRAE Work and WorkBuddy follow a different path: the multi-model aggregation route. WorkBuddy supports switching between multiple models, including Hunyuan, GLM, MiniMax, Kimi, and DeepSeek, and offers an Auto mode that selects the model based on the task. TRAE Work also currently integrates various models and provides an Auto mode.
Even following the aggregation route, the two products do not have the exact same set of model cards. Based on the current model selection available in the products, WorkBuddy's default model pool does not include Alibaba's Qianwen or ByteDance's Seed. TRAE Work, meanwhile, does not include Tencent's Hunyuan in its visible model selection but does include the 3.7 version of Qianwen.
This difference is not yet sufficient to suggest fixed camps among large model companies, as models like Seed can also be integrated into third-party agents via APIs. The boundaries of the model pool hint at the competitive tension between Tencent and ByteDance: on one hand, WorkBuddy and TRAE Work want to incorporate more models to become universal workbenches; on the other hand, both Tencent and ByteDance own their own proprietary models (Hunyuan and Seed, respectively), which are themselves competing for call volume and user usage.
This raises a new question: As office agents begin to take on more real-world tasks, will they become new channels for model distribution? Will the battle for the desktop entry point further evolve into a battle for model call volume?
Is Routing the Next Challenge?
When an agent has multiple models at its disposal, how to orchestrate them becomes an unavoidable problem for the aggregation route. Behind this is the growing differentiation of model capabilities. As models continue to iterate, the strengths of different models in areas like coding, complex reasoning, long-text processing, search, response speed, and cost are becoming more distinct.
This difference is amplified in agent scenarios. A complex task often requires multiple consecutive model calls, and the choice of model at each step directly impacts the task's completion quality, time, and cost. Although routing happens in the background, its impact on user experience is increasingly direct. In an ideal scenario, the user simply proposes a task, and the backend automatically selects the model: simple information extraction is handled by a faster, cheaper model, while complex reasoning and coding tasks are handled by a more capable one.
For individual users, this translates into wait times and point consumption; for enterprises, it becomes a matter of computing cost. Finding the balance between task effectiveness and computational cost is the most direct user value of routing. If aggregated agents like WorkBuddy and TRAE Work eventually accumulate enough users, routing could also create another value: a data flywheel. Each real-world task leaves feedback on model performance, such as which model is more successful for which task, how long it takes, how many tokens are consumed, and in which scenarios it fails.
Subject to privacy, authorization, and data governance requirements, this feedback can be used to optimize routing strategies. This could also change the relationship between model companies and agent platforms. In the past, large model companies competed for users and developers to actively call their models. As more tasks are completed by agents, users may gradually delegate model selection to the backend. For model companies, being part of the routing layer means being part of more real-world tasks. If a model develops a clear advantage in specific tasks like coding, search, or reasoning, it could gain more call volume.
Competition between models may extend from vying for users' active choice to vying for selection by the routing system. Users need routing to manage model selection and computing costs, while model companies need this layer to reach more real-world tasks. The routing layer thus sits between model supply and user demand. The capital markets are also beginning to recognize the value of aggregation platforms. In May, OpenRouter, the world's largest model aggregation platform, completed a $113 million Series B funding round, reaching a post-money valuation of $1.3 billion.
However, whether this commercial value can be truly realized depends on the accuracy of routing. Currently, there is no industry-wide consensus on this. The differing approaches within Tencent's own research teams illustrate this divergence. The DecoR research, a collaboration between Tencent Hunyuan and the University of Chinese Academy of Sciences, focuses on single requests, breaking down a query into capability needs like skills, knowledge, and difficulty, and then searching for the best matching model based on historical performance. Another study, a collaboration between Tencent's WeChat team and Shanghai Jiao Tong University, takes a longer-term view, arguing that a model that performs best on a single step may not lead to the optimal outcome for a full task sequence, and that model switching can impact cache reuse and cumulative costs.
One focuses on "which model should handle this single request," while the other begins to consider "what the final outcome of a series of model selections will be." When exactly to switch models and how often remain questions without standard answers. But for aggregated agents, the commercial question is quite specific: whoever can more accurately assign real-world tasks to the right model while reducing computing costs will have a better chance of turning multi-model capabilities into a true product advantage.
Comments