A new wave of competition among Chinese tech giants is heating up on the corporate desktop.
On July 30th, ByteDance initiated a major organizational restructuring for its AI business. The Feishu (Lark) product team was integrated with the Doubao product team to form a new Doubao product team. Separately, the Feishu GTM (Go-to-Market) team was merged with the Volcano Engine team to create a new To-B organization called the "Creativity Service Platform," tasked with unifying the marketing, sales, and customer service for both MaaS and SaaS products.
Following this reorganization, the enterprise version of Doubao, developed with deep involvement from the Feishu team, has begun internal testing with select Feishu clients. Users can now access office capabilities like Feishu documents, spreadsheets, meetings, and group chats without needing any additional deployment.
Just a week prior, Alibaba announced the integration of its internal products—QoderWork, Wukong, and MuleRun—to launch "Qianwen Office," placing the product under the management of DingTalk.
Ten days before that, Tencent also completed a round of adjustments, merging the entire QClaw product center into the team responsible for WorkBuddy, thereby further strengthening its desktop office Agent strategy.
This, however, does not signal an end to the internal "horse racing" mechanism within these large companies.
Beyond WorkBuddy, Tencent also has Marvis, which can access the operating system and local files, as well as "Da Yuan," an enterprise WeChat Agent currently in internal testing. Similarly, ByteDance, alongside the Doubao enterprise version, has retained another desktop office Agent, TRAE Work.
Therefore, these recent adjustments seem more like a move by the tech giants to consolidate overlapping projects, centralize underlying capabilities, and pool commercialization resources, rather than a definitive selection of a single product form. Internal competition continues, and it will be the market that ultimately determines which product finds the most frequent use-case scenarios and a stable product boundary first.
Three Titans Reorganize Simultaneously for Desktop Agent Dominance
Within just half a month, Alibaba, Tencent, and ByteDance have all reorganized around the desktop office Agent. The driving force behind this may be the battle for a new kind of data that has never been captured at scale before: real-world workflow data.
In the past, large language models primarily learned from internet text and user dialogues. But when an Agent starts participating in document writing, data analysis, meeting coordination, and approval workflows, it gains access to the complete process of how a task is broken down, executed, revised, and finally delivered.
This marks the beginning of a new war in the summer of 2026, centered on "how AI learns to work."
Office Agents Take the Stage
It is no coincidence that three major companies have chosen similar paths almost simultaneously. A key reason is likely that WorkBuddy was the first to validate the user growth potential of desktop office Agents, forcing competitors to stop waiting on the sidelines.
According to data from Analysys, in June of this year, Tencent's WorkBuddy desktop client had 20.97 million visits, ranking first among domestic AI-native office Agent products and surpassing the combined visits of the second and third-place products.
This has transformed the desktop office Agent from an internal exploration project within tech giants into a competition with a clear user scale benchmark.
Concurrently, the AI office product landscape is entering a phase of "convergence." Over the past year, it was common to see multiple teams and product lines within a single company exploring in parallel. Alibaba's QoderWork, Wukong, and MuleRun; Tencent's QClaw and WorkBuddy; and ByteDance's Doubao and Feishu all sought the ideal product form for office Agents from different angles.
As the market enters a phase of "calculating returns," the cost of maintaining multiple similar product lines has increased. Resources for models, R&D, computing power, sales, and customer relationships need to be refocused on products with the highest potential for user scale and revenue. Alibaba's integration of three products into Qianwen Office and Tencent's merger of the QClaw team into WorkBuddy are clear signs of this convergence.
A more pragmatic issue than product convergence comes from the enterprise market itself. Selling a standalone office Agent to a business requires the company to learn about a new product, evaluate security and permissions, complete the procurement and deployment process, and then train employees to change their work habits. This process has a high educational cost and a long sales cycle.
A faster path to market is to have the office Agent directly integrate into the collaboration systems that enterprises are already using. A salesperson from Alibaba Cloud noted that integrating Qianwen Office into DingTalk makes it much easier to sell to enterprise clients.
The logic is straightforward. DingTalk already has mature organizational structures, permission systems, and office scenarios. Once Qianwen Office is integrated, salespeople no longer need to market a standalone AI product from scratch. Instead, they can introduce the Agent's capabilities around a company's existing workflows for meetings, documents, approvals, and collaboration, making it easier for clients to understand and more likely to make a purchase.
Currently, Qianwen Office is sold as an independent product without a dedicated sales team. However, as the product is newly launched, internal sales strategies are constantly being adjusted based on market feedback.
ByteDance's recent integration of its product and sales systems reflects a similar strategy. After the Feishu product team was merged into Doubao, the Doubao enterprise version can directly leverage Feishu's capabilities in documents, spreadsheets, meetings, and group chats. The merger of the Feishu GTM team with Volcano Engine allows ByteDance to promote the Doubao enterprise version using existing enterprise customer relationships, thereby reducing the cost of re-acquiring customers and delivering standalone solutions.
The outcome of this competition is far from certain, but one company that may feel the immediate impact is Kingsoft Office. With WorkBuddy, Qianwen Office, and the Doubao enterprise version entering the market in rapid succession, Kingsoft Office's competitive landscape has changed. Its rivals possess large language models, cloud services, user traffic, and enterprise sales systems, and can accelerate customer acquisition through free or low-cost pricing strategies.
On July 21st, Goldman Sachs downgraded Kingsoft Office's rating from "Neutral" to "Sell," citing the potential for WPS AI's user adoption and payment rates to take longer to improve. The rapid expansion of the AI office market and the more attractive pricing of competing products are also seen as pressures on its subscription business. The battle for the desktop Agent among tech giants has just begun, but traditional office software is already feeling the heat.
The Real Prize: Capturing Authentic Workflows
For Alibaba, Tencent, and ByteDance, the desktop office Agent holds an even deeper value: allowing AI models to observe the complete process of how a task is completed. Historically, the data available to large models has come mainly from internet text and user dialogues. A model knows what question a user asked and what answer it generated, but it struggles to know if that answer truly solved the problem. It certainly doesn't know how a real task is accomplished step-by-step.
Desktop office Agents change this. When an Agent operates a computer, opens a browser, accesses documents, analyzes spreadsheets, creates presentations, and collaborates with a user to complete a task, it records not just a few rounds of dialogue but a complete task execution trajectory (Agentic Trajectory).
For example, if a user asks an Agent to complete a piece of industry research, the Agent would search for information, read PDFs, organize data, and generate a draft. The user might then adjust the structure, add new materials, and ask the Agent to revise it, leading to a final deliverable. This entire workflow is becoming a new form of data for the Agent era.
A recent Agent evaluation survey published by Springer notes that unlike traditional machine learning, which primarily processes "input-output" data, Agent systems must continuously interact with their environment. Therefore, evaluating an Agent requires more than just looking at the final result; the complete trajectory must be recorded. The paper argues that a full trajectory includes not just the final state but also intermediate reasoning, tool calls, environmental feedback, and error correction. These trajectories truly document how a job is done, providing much richer data than the final answer alone.
This type of data has become a hotly contested new resource in the AI industry. A manager at an AI platform company in Shanghai revealed that some model companies have approached them for collaboration, specifically to obtain long-horizon task-specific data from their industry. These companies are paying high prices for such rare datasets.
This hunger for real-world workflows isn't limited to China. Overseas model giants like OpenAI and Anthropic are also investing heavily in sending Field Deployment Engineers (FDEs) directly into enterprises. In May, Anthropic announced the formation of an enterprise AI services company with institutions like Blackstone, Hellman & Friedman, and Goldman Sachs. Under this plan, Anthropic engineers will work with the new company's team within client enterprises to find business processes where Claude can be integrated and develop systems around existing workflows.
A week later, OpenAI established OpenAI Deployment Company. This new company, majority-owned by OpenAI, has an initial investment of over $4 billion and plans to acquire the AI consulting and engineering firm Tomoro, bringing in its approximately 150 FDEs and deployment experts. The prevailing market view at the time was that Anthropic and OpenAI were following Palantir's playbook to shore up their enterprise services and project delivery capabilities.
However, the head of the Shanghai AI company believes the deeper purpose for these two giants goes further. They see it as a way to deploy FDEs into enterprises to understand real workflows, compete for workflow data, and learn the underlying methodologies. Frankly, most current AI models only deliver results and are still quite lacking in understanding a business's operational logic, preferences, and other specific information.
If internet text shaped the knowledge capabilities of models like ChatGPT, then real-world workflow data may be what shapes the execution capabilities of the next generation of Agents.
The Next Stage of Evolution
At this stage, there is little discernible difference between WorkBuddy, Qianwen Office, and ByteDance's related products on general office tasks. The ability to search for information, read files, generate documents, analyze spreadsheets, create presentations, and call upon browsers and other software to complete tasks has become the baseline for desktop office Agents. These features will likely become even more similar as underlying models and Agent frameworks continue to iterate.
In past internet competitions, when product capabilities converged, price subsidies, traffic portals, and user scale often became decisive factors. However, this strategy is difficult to replicate in the desktop office Agent market. During the Spring Festival, several internet companies used activities like giving away red packets, offering freebies, and ordering bubble tea to attract users to download and use their large model products. Subsidies can quickly lower the barrier to a first-time experience, but paying for a business tool is different.
An enterprise pays for an office Agent to get reliable, stable work done. The Agent must understand the company's organizational structure, business rules, and permission boundaries. It needs to know what materials to read, which systems to call upon, which steps to follow for a task, and where to pause and ask for confirmation. Subsidies can generate trial users, but they can't keep a poorly performing Agent that doesn't understand the business in a company's core processes. The market will ultimately be decided by whether the product can consistently get the job done.
Currently, the three major players have established different product strategies. Qianwen Office is a standalone desktop client but is also bidirectionally integrated with DingTalk. Users can initiate tasks from DingTalk, and after the Agent completes them, the results are returned to the original chat and collaboration flow. Beyond DingTalk, Qianwen Office can also connect to external platforms like Feishu to read their documents with authorization. This suggests Alibaba wants Qianwen Office to cover as much of an enterprise's existing software and data environment as possible, though this openness is mainly at the software and data connection layer, with Qianwen Office primarily using the Qwen model.
ByteDance, on the other hand, has kept two product lines. One is the Doubao-Feishu combination, which enters enterprise workflows through chat, documents, meetings, and organizational relations, primarily using its own model. The other is TRAE Work, which started as a coding tool and was renamed from TRAE SOLO in June. Its product scope has expanded to include research, writing, data analysis, strategy development, and cross-team collaboration. Unlike Doubao, TRAE offers multiple model options for users to choose from based on the task.
Tencent, too, has more than just WorkBuddy. WorkBuddy itself is a standalone desktop workspace, emphasizing cross-software invocation, task execution, and automation. Unlike Qianwen Office and Doubao, it highlights multi-model aggregation, allowing users to switch between different models and invoke software like WPS and Tencent's suite to complete tasks. Then there's "Da Yuan," the enterprise WeChat AI Agent that entered internal testing in June. It grows directly within enterprise WeChat and can understand user requests by integrating work data from group chats, documents, meetings, emails, and calendars, making it a more native Agent within the enterprise collaboration environment. Overall, the internal "horse races" within these tech giants are far from over.
Divergent Strategic Paths
Looking at the big picture, two distinct strategic paths are emerging for desktop office Agents. The first is the proprietary model path, represented by Qianwen Office and Doubao. Here, the office Agent is built on the company's own model, and the model, office software, cloud services, and enterprise data are all housed within a single, integrated system.
The second is the multi-model aggregation path, represented by WorkBuddy and TRAE. On this path, the platform doesn't tie its capabilities to a single model. Instead, it offers multiple model options and connects them to files, software, and task processes through a unified workspace.
These different routes also imply different ways of building a competitive moat. For the proprietary model path, the primary focus is on model training and post-training. The model's ability to understand longer, more complex tasks and to process documents, spreadsheets, code, and multimodal information more accurately sets the upper limit of the office Agent's capabilities. The real-world tasks and user feedback generated in office scenarios can, in turn, help the model iterate further.
For the multi-model aggregation path, the underlying models are not proprietary. The same model can be accessed by multiple competitors, so simply adding more models is unlikely to create a long-term advantage. The true differentiator lies in what's outside the model: the harness. The harness is responsible for organizing context, connecting tools, saving task state, managing permissions, handling errors, and converting the model's reasoning abilities into a practical, runnable workflow.
The billions of yuan spent on Spring Festival subsidies for large models were a battle for the first-time usage of C-end consumers. The battle for office Agents, however, is a battle for the experience of getting work done. The key difference lies in where that experience ultimately ends up: is it incorporated back into the model itself, or is it stored within the execution system that sits outside the model?
Comments