文 | 深流研究所，作者 | 之丰
AI办公大战打到现在，大家开始齐刷刷走腾讯的路了。
此前，用户打开哪家的办公Agent，就是用哪家的模型。腾讯WorkBuddy是少数派，在今年3月上线之初，它就向用户开放了DeepSeek、GLM、Kimi等三方模型的选择。
10月以来，WorkBuddy又陆续接入阶跃星辰和小米的旗舰模型，几乎涵盖所有国内主流大模型。
现在，其他AI办公产品开始跟上WorkBuddy的多模型策略了。上周五，豆包工作开始向企业客户开放三方大模型服务，首批接入了Kimi、GLM、DeepSeek等系列模型。就在同一天，小米MiMo Desktop也宣布会员用户可以选择更多模型，包括DeepSeek系列。
为什么AI办公产品都开始把对手的模型请进自家？
任务面前，没有永远的单点领先
多模型这条路，WorkBuddy走了半年多后，越来越多AI产品开始跟着走了。
10月初，马斯克在X上宣布，Grok Bot将可以调用Claude、Midjourney、Suno等第三方模型。几天后，谷歌发布面向办公场景的Gemini Agent，也开放了对Claude等第三方模型的支持。
微软在10月底发布Copilot Harness的文章开篇，就指出了多模型策略的必要性：“If you want to compete, you need to be multi-model.”在微软看来，多模型已经成为AI公司参与竞争的入场券。
背后的原因并不复杂。AI进入办公场景后，是以任务执行为中心，用户所需的智能供给也就变得多元。
即使是在同一家企业里，行政、销售、财务、设计、产品经理等不同岗位的任务也相差甚大。况且，一项工作任务可能就包含资料检索、数据分析、内容生产、视频生成等多个环节，这对模型的推理、写作、多模态等能力提出了不同要求。
与此同时，如今模型厂商在不同能力维度上持续竞争，迭代速度加快。这个阶段，某家模型可能在编程任务上表现突出，另一家可能在成本效率上更有优势，而随着新模型发布，原先的相对优势又会被重新洗牌。
能力上的领先像潮水，涨落之间，没有哪一家能在岸上站定不动，谁也没法保证自家模型能在所有任务上永远领先。
所以，马斯克和谷歌在解释为何提供三方模型时，说法都很一致：任务不同，所选的模型也理应不同。
让任务匹配模型，便成为多模型策略的价值所在。Google Cloud的CEO在10月初时公开表示，“最适合任务的模型并不总是最大的”。让模型与任务的匹配，可以提高复杂任务的准确率，并降低简单任务的成本。
尤其是在办公场景下，用户的任务记录、工作流和上下文等，往往都沉淀在一个Agent中。开放模型选择后，用户能根据任务进行质量、速度与成本的优化，无需为单一模型妥协；也不必因为模型偏好改变，重新迁移整套工作环境。
对于各家巨头来说，留住用户的整套工作环境，比坚持只使用自家模型更有商业价值。
模型隐入后台，Agent入口更加重要
以任务为核心的需求，正在改写模型能力与Agent产品之间的关系和位置。
对办公Agent来说，模型是完成任务的底层智能供给。Agent需要整合这些能力，让它们能够在具体的工作流程中发挥作用。
如果一款办公Agent无法提供用户完成任务所需的更丰富、更可靠的智能能力，用户就可能转向其他产品。豆包工作首先向企业用户开放第三方模型，多半也出于这层考虑。
与此同时，模型能力也需要寻找更多进入真实任务的渠道。
在过去，大模型厂商通常通过自家的Chatbot或API向用户和开发者提供能力。随着办公Agent的发展，模型能力抵达用户的路径正在变得更加多样。
据易观分析数据，2026年9月，中国桌面端AI原生办公智能体市场合计访问量已突破3亿次，同比2026年6月增长近5倍，其中排在首位的WorkBuddy9月月访问量就达8534万次。
办公Agent正在形成规模化的用户触达渠道。所以，模型厂商也在开发自己的Agent，比如DeepSeek9月底就上线了DeepSeek Harness。
随着模型能力逐渐变成底层供给，Agent作为用户使用模型能力的入口，其价值变得愈发重要。
据腾讯披露，WorkBuddy自今年3月上线以来，已经完成90亿次任务请求，企业版进入50多个行业，覆盖50%的中国企业500强和60%的中国企业100强。所以，最近DeepSeek发布的V4.1 Flash新版本模型，就特别针对WorkBuddy等平台框架进行了优化。小米在宣布旗下Agent产品接入DeepSeek的同时，它的大模型也接进了WorkBuddy里。
向外调度模型，对内重组自身
接入更多的模型，只是Agent能力建设的一部分。
不同模型接入之后，Agent还需要解决工具调用、上下文传递、执行控制和结果验证等问题。只有这些环节能够协同工作，多模型能力才能转化为真正的任务完成能力。
这又牵出AI办公竞争的另一面：公司要如何围绕任务协同，重组自身的能力与资源。
上述易观分析数据里前三名的办公Agent，分别来自腾讯、阿里和字节。这些成熟的科技大厂原本就拥有丰富的相对独立的产品线，各自积累了用户入口、数据、工具、账号体系和产品团队等。
而在真实的办公场景中，用户交给Agent的往往是一项跨越多个产品、数据源和操作环节的任务。Agent需要检索分散在不同系统中的信息，调用相关产品的能力，再把执行结果连接起来。如果各个产品仍然只能独立提供功能，用户就需要在它们之间手动切换、搬运资料和重复交代背景。
办公Agent要真正完成任务，就需要将这些原本分散的能力和资源重新组织起来。近期腾讯WorkBuddy和企业微信全面打通，以及上周刚发布的Buddy One，都是围绕Agent任务协同再造自身。Buddy One把WorkBuddy和腾讯文档、腾讯会议、腾讯网盘、ima接到了一起，用户授权之后，文档、会议记录和知识库可以一并检索，账号和积分权益也统一了。
其中，新推出的腾讯网盘为WorkBuddy套件提供统一存储，让Agent工作中积累的资料和生成的成果能够保存下来，并在后续任务中按权限继续调用。用户切换设备或者应用，工作也不必从头来过。
不同产品提供的功能和内容，由此能够围绕同一项任务协同起来。腾讯的内部整合还延伸到了云基础设施。据腾讯介绍，腾讯云Agent Runtime针对Agent长时间运行、状态保存、休眠与恢复等特点，调整了基础设施，让任务能够持续执行。
上个月刚发布AI办公产品的谷歌也在进行类似的探索。Gemini Agent能够连接Gmail、Drive、Docs、Slides、Sheets、Chat和Calendar等Workspace应用，支持跨应用执行任务，并利用相关工作资料和上下文提供服务。它也可以根据任务需要调用不同模型、工具和子Agent，包括Claude等第三方模型。
此外，腾讯和谷歌的整合没有停留在自身的产品体系内。比如，WorkBuddy可以通过开放能力接入飞书等第三方办公产品。Gemini Agent也支持连接Confluence、Microsoft Office、Teams、Slack等外部办公工具。办公Agent如果只能调用自家产品，任务执行仍然会受到生态边界的限制。
腾讯和谷歌，都在尝试让旗下的AI办公产品承担起一个双向调度枢纽的角色。向外接入三方模型和外部生态，根据任务调用不同的能力；对内调度产品、工具和上下文等，让这些资源围绕任务协同。
当多模型策略逐渐成为共同选择，AI办公的竞争也将进一步深入到任务执行过程。模型可以来自不同厂商，但用户把任务交给一个Agent之后，真正影响体验的，是它能否理解目标、找到合适的能力、组织执行过程，并在出现问题时及时调整，最终交付可靠的结果。这份统一调度的功夫，可没法从别家轻易学过来。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Author, Institute of Deepstream Studies
So far, the AI office war has begun to clear the Tencent road.
Previously, the user opened the office Agent, which model it used. TencentWorkBuddy, a minority, opened the choice of three models, DeepSeek, GLM and Kimi, to users at the beginning of March.
Since October, WorkBuddy has continued to access the flagship models of skyrocketing stars and millet, covering almost all of the country ' s mainstream LLM.
Now, other AI office products are following WorldBuddy's multi-model strategy. Last Friday, Doubao began to open three-way LLM services to business clients, the first of which involved a series of models such as Kimi, GLM and DeepSeek. On the same day, MiMo Desktop also announced that member users could choose more models, including the DeepSeek series.
Why does the AI office product start inviting the opponent's model home?
There's no single-point lead in front of the mission.
After more than six months on the multi-model road, a growing number of AI products began to follow.
In early October, Mask announced on X that Grok Bot would have access to third-party models such as Claude, Midjourney and Suno. A few days later, Gemini Agent, a Google publish-oriented office scene, also opened up support for third-party models such as Claude.
At the end of October, Microsoft opened its article by pointing to the need for a multi-model strategy: “If you want to compete, you need to be multi-model.” In Microsoft's view, the multi-model has become a competitive entry ticket for AI.
The reasons behind this are not complex. When AI enters the office scene, it is mission-driven, and the user ' s intellectual requirements are diversified.
Even in the same enterprise, the tasks of the different positions of administration, sales, finance, design and product manager vary considerably. Moreover, a task may involve a number of components, including information retrieval, data analysis, content production, video production, etc., which may require different capabilities in terms of reasoning, writing, polymorphism, etc.
At the same time, there is now continuous competition among model producers on different dimensions of capacity, accelerating the iterative pace. At this stage, one model may be prominent in programming tasks, another may be more cost-effective, and with the new model publish, the original comparative advantage will be redoed.
No one can stand on shore, no one can guarantee that the home model will always lead all tasks.
So, in explaining why the three-way model was offered, Musque and Google were all in agreement: different tasks, and the model chosen should be different.
Making task matching models is the value of multi-model strategies. Google Cloud's CEO publicly stated in early October that “the best model for the mission is not always the largest”. Matching the model with the mission will increase the accuracy of complex missions and reduce the cost of simple missions.
Especially in the office setting, user mission records, workflow and context, among others, tend to sink in an Agent. Following the selection of an open model, the user is able to optimize the quality, speed and cost of the mission without having to compromise for a single model; nor is it necessary to relocate the entire work environment because of a change in model preferences.
For giants, the entire work environment in which users are retained is more commercially valuable than the insistence on using only home-based models.
The model is hidden backstage and the Agent entrance is more important.
Mission-centred demand is recasting the relationship and location between modelling capabilities and Agent products.
For Agent, the model is the bottom supply of intelligence for the mission. Agent needs to integrate these capabilities so that they can play a role in the specific workflow process.
If one office, Agent, does not provide the richer and more reliable intellectual capabilities that users need to fulfil their mandates, users may turn to other products. Doubao is the first to open up third-party models to business users, mostly for this consideration.
At the same time, model capabilities need to find more access to real task.
In the past, LLM manufacturers usually provided capacity to users and developers through their own Chatbot or API. With the development of office Agent, model capacity is becoming more diversified to reach users.
According to easy-to-see analysis data, in September 2026 the total number of visits to the AI agent market in China ' s desktop-end primary AI office exceeded 300 million, nearly five times the same rate as in June 2026, of which the number of visits in the first WorldBuddy in September was 8.5340 million.
The office of Agent is forming a scalable user access channel. So modelers are also developing their own Agent, like DeepSeek, who went online at the end of September.
As model capabilities gradually become a bottom supply, Agent ' s value becomes even more important as an entry point to users ' ability to use models.
According to Tencent, WorldBuddy has completed 9 billion mission requests since it went online in March of this year, with the enterprise version entering more than 50 industries, covering over 50 per cent of Chinese enterprises, over 500 and over 60 per cent of Chinese enterprises. As a result, the latest version of the V4.1 Flash model of DeepSeekpublish has been optimized for platform frameworks such as WorkBuddy. While Mi announced that the Agent product was connected to DeepSeek, its LLM was also connected to WorkBuddy.
Outward scheduling model, reorganizing itself internally
Access to more models is only part of Agent capacity-building.
After access to different models, Agent also needed to address the issues of tool call, context transmission, executive control and validation of results. Only if these links work together can the multi-model capability be translated into a real mission completion capability.
This, in turn, leads to the other side of AI's office competition: how companies can reorganize their capacities and resources around the task.
The first three of these visual analysis data, Agent, are from Tencent, Ali and byte. These mature technology giants are already rich in relatively separate product lines, each accumulating user accesses, data, tools, account systems and product teams.
And in a real office scene, it is often the task of the user to Agent to cross multiple products, data sources and operations. Agent needs to retrieve information that is dispersed across different systems, and the ability to access relevant products, linking the results. If products are to remain functionally independent, users need to manually switch, move information and repeat the context between them.
To truly fulfil its mandate, Agent needs to reorganize these previously dispersed capacities and resources. Recently, TencentWorkBuddy and the company WeChat, as well as Buddy One, just last week ' s Z94Z, were working together to recreate themselves around the Agent mission. Buddy One has brought together WorkBuddy and Tencent files, Tencent meetings, Tencent webpas, ima, which, with the authorization of the user, will allow the documents, meeting records and knowledge base to be retrieved together, and account numbers and points of interest to be harmonized.
Among them, the newly launched Tencent webpad provides a unified storage of the WorkBuddy suite, allowing for the preservation of information accumulated and the results generated during Agent ' s work and continued to be called on permission for subsequent missions. The user switches the device or application, and the job does not have to start from scratch.
The functions and content provided by different products can thus be aligned around the same task. Internal integration of Tencent also extends to cloud infrastructure. According to Tencent, the Tencent cloud Agent Runtme has adapted its infrastructure to enable the mission to be carried out on an ongoing basis, in response to the features of Agent ' s long operation, state preservation, hibernation and recovery.
Last month, a similar search was carried out in Google for publish AI office products. Gemini Agent is able to connect to the Workspace applications of Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, to support cross-application missions and to provide services using relevant work materials and context. It can also use different models, tools and sub-Agents, including third-party models such as Claude, according to mission needs.
Moreover, the integration of Tencent and Google did not remain within their own product systems. For example, WorkBuddy can access third-party office products such as flying books through open capacity. Gemini Agent also supports the connection to external office tools such as Confluence, Microsoft Office, Teams and Slack. The implementation of the mandate of the office of Agent remains constrained by ecological boundaries if it can only call on its own products.
Both Tencent and Google are trying to make the AI office product a two-way dispatch hub. Outward access to the tripartite model and the external ecology, with different capabilities deployed according to the mission; internal movement of products, tools and context, etc., to bring these resources around the mission.
As multi-model strategies become a common option, the competition for AI offices will go further into the task implementation process. The model can come from different manufacturers, but the real impact experience after the user has given the task to an Agent is its ability to understand the objectives, find the right capabilities, organize the process of implementation and adjust in a timely manner when problems arise, ultimately delivering reliable results. It's not easy to learn from others.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.