Summary of Key Points
Recently, the competition among large tech companies in the field of AI-powered office assistants has become more practical and tangible. ByteDance’s TRAE Work and Tencent’s WorkBuddy have even started competing for advertising space in elevator halls of residential buildings. Behind this competition, the companies have adopted two different technical approaches: one focuses on developing their own models (such as Alibaba’s Qianwen Office and ByteDance’s DouBao Office, which rely on their own large-scale models), while the other aims to integrate models from multiple vendors (like TRAE Work and WorkBuddy, which combine models from various providers in a single platform). Products that follow the integration approach are now facing a crucial challenge: how to automatically select the most suitable model for tasks in the backend (a process known as “routing”). This not only affects user experience but could also become a new battleground for model vendors.
The Hidden Battle in Elevators: Big Tech Firms Competing for Users
The advertisements in your neighborhood’s elevators may reflect the competitive strategies of these companies. Not long ago, an elevator in a Chaoyang district of Beijing displayed ads for Tencent’s WorkBuddy; now it shows ByteDance’s TRAE Work. This is no coincidence—both companies are trying to attract the attention of ordinary office workers.
TRAE Work was originally designed as an AI programming tool (called TRAE SOLO) and was upgraded in June this year to cover a range of general office tasks such as writing reports, data analysis, and cross-team collaboration, making it accessible to everyone rather than just programmers. Within ByteDance, there is another approach: DouBao Office, which integrates with Lark and uses ByteDance’s own Seed model. By launching TRAE Work first to compete with WorkBuddy, ByteDance is adopting a dual-pronged strategy, but it also needs to address how these products can coexist within the company.
The Two Approaches to Office Assistants: Developing Own Models vs. Integrating Multiple Models
Large tech companies are taking two different approaches to creating office assistants:
- Self-developed Model Approach: Companies like Alibaba’s Qianwen Office and ByteDance’s DouBao Office rely on their own models (Qianwen and Seed, respectively, along with Lark’s workflow). This approach allows the products to perform more tasks, but it requires continuous investment in model training.
- Multi-model Integration Approach: TRAE Work and WorkBuddy combine models from various vendors (e.g., WorkBuddy uses Tencent’s Hunyuan and GLM, while TRAE Work uses Alibaba’s Qianwen 3.7). Users don’t need to switch models manually; the backend automatically makes the selection. This approach offers a wider range of tasks, but selecting the right model remains a significant challenge.
Interestingly, the model pools of these integrated products are mutually exclusive: WorkBuddy doesn’t include ByteDance’s Seed model, and TRAE Work doesn’t include Tencent’s Hunyuan model—after all, they are competitors.
The Challenge of Routing: The Behind-the-Scenes Decision-Making for Integrated Agents
The biggest challenge for integrated agents is routing—the process of automatically selecting the most suitable model for a task. For simple tasks like extracting information, a fast and cost-effective model will suffice; for complex tasks such as financial analysis, a more powerful model is needed. The quality of this routing directly affects user experience: the right model can complete tasks quickly and efficiently, while the wrong choice can result in slower performance or wasted resources (either for individual users or for enterprises).
More importantly, the more users an integrated agent is used, the more data the backend collects about model performance (e.g., which model performs well on coding tasks). This data creates a positive cycle where more usage leads to better models, which in turn attracts more users (a phenomenon known as a “data flywheel”). However, there is no consensus in the industry yet on the best approach. Even within Tencent, there are different research directions: some focus on selecting the right model for a single task, while others consider the overall efficiency of consecutive tasks (e.g., whether choosing one model for one step and another for the next step is more cost-effective).
A New Battleground for Model Vendors: Routing as a Traffic Gateway?
Previously, model vendors directly competed for users by encouraging them to choose their models. Now, integrated agents could become new distribution channels, with the routing mechanism determining which models get the most usage. For example, if a model performs exceptionally well on coding tasks, it will be frequently selected, increasing its usage and generating more revenue for the vendor. This is similar to the “recommendation systems” in e-commerce platforms, where vendors compete for favorable routing decisions.
The capital market has recognized this potential: OpenRouter, the world’s largest model integration platform, raised $113 million in May this year with a valuation of $1.3 billion. However, success will depend on the accuracy of routing—only by efficiently assigning tasks to the right models while reducing costs can vendors emerge as winners.
In Conclusion
The competition among office assistants is ostensibly about advertising in elevators and attracting users, but it’s actually about competing technical approaches, with the ultimate goal of making AI more capable of completing tasks efficiently. For ordinary users, in the future, they may not need to worry about choosing a specific model; an office assistant could automatically handle complex tasks for them as long as the big tech companies solve the routing problem first.