三足鼎立:2026年中国AI办公智能体时代Tripod Stand: The 2026 Chinese Era of AI Office Agents


AI free token/AI免费积分

腾讯orkbuddy注册即可得2000积分,可以干不少活。下载链接:https://www.workbuddy.cn/

百度Dumate智能体,也是工作搭配好工具:邀请码及下载地址:百度 Dumate邀请码: 3V3JPSD

字节TraeWork免费4500积分,点击下载字节跳动TraeWork免费4500积分

 

2026年夏天,中国AI办公赛道迎来真正的拐点。

易观数据显示,国内17款主流桌面端AI原生办公智能体,2026年6月合计月访问量突破6000万次——而这一数字在3月仅为2000万。短短三个月,用户量翻了三倍。办公智能体正从"尝鲜工具"蜕变为职场人的日常生产工具。

在这场入口之争中,腾讯WorkBuddy、百度DuMate、字节Trae Work三款产品,分别代表了三种截然不同的技术路径与生态打法。它们背后是三家互联网巨头对"AI如何重塑办公"这一命题的不同回答。


一、腾讯WorkBuddy:从"一句话"到"交付"的全场景工作台

核心定位:全场景职场AI智能体桌面工作台

WorkBuddy于2026年3月正式发布,是腾讯旗下定位最明确的AI办公智能体产品。据腾讯2026年Q1财报,按日活跃账户数计,WorkBuddy已成为中国最受欢迎的效率AI智能体服务。

关键数据: - 2026年6月月访问量2097万次,位居国内AI办公智能体市场第一 - 超过第二、第三名月访问量总和 - 7月18日WAIC 2026上发布独立APP,成为业界首个同时登陆iOS、Android、鸿蒙三端的通用智能体应用

核心能力:

WorkBuddy最大的特点是"全流程自主执行能力"——它不是一个聊天机器人,而是一个能直接帮你干活的数字员工。面对"全球人口结构变迁可视化分析"这类复合型需求,WorkBuddy能像资深员工一样自主规划步骤、读取分析本地文件及在线资料、调取绘图技能并自动完成排版与看板开发,直接交付专业级成品。

  • 多模型灵活切换: 全面接入腾讯混元Hy3、MiniMax、Kimi、DeepSeek、GLM等主流大模型,更支持通过API接入私有模型。混元Hy3上线后,选择该模型的用户占比已达60%以上
  • 复杂任务自主拆解与闭环: 不仅能"理解"需求,更擅长"执行"——自主规划步骤、调用工具、循环试错直到交付
  • 本地文件自主管理: 获得授权后可直接读写本地文件夹,秒级完成批量重命名、跨表格数据提取、文件自动归档等工作
  • 生态连接: 打通腾讯文档、企业微信等腾讯生态;上线技能中心、专家模式、自动化定时任务、资料库等能力
  • 多端协同: 支持PC、APP、微信小程序多端,可远程下发并执行任务

差异化优势: 腾讯完整的社交与办公生态(微信、企业微信、腾讯文档、腾讯会议)为WorkBuddy提供了天然的连接优势。7月6日上线的混元Hy3模型以295B总参数、21B激活参数实现"小尺寸比肩开源旗舰"的智能水平,实测可胜任90%的Agent任务,成为WorkBuddy的强力引擎。


二、百度DuMate:从"想法"到"结果"的超级执行力

核心定位:真干活,用搭子

百度搭子DuMate于2026年3月发布桌面版,5月13日在Create2026百度AI开发者大会上推出移动端App。DuMate在多个国际权威Agent Benchmark评测中达到SOTA水平,7月入选WAIC"镇馆之宝",成为唯一获此殊荣的通用智能体产品。

核心能力:

DuMate的核心卖点是"超级执行力"——只需一句话,就能实现从"想法"到"结果"的自动执行。它将百度生态的核心产品能力整合为可随时调用的内置技能:

  • 百度生态深度整合: 将百度AI搜索、秒哒、伐谋、百科等能力集成为内置Skill,通过统一端口面向用户。相比需要调用三方工具的智能体,百度从基础模型到垂直智能体的完整生态,让工具调用更加丝滑
  • 长程任务执行与主动决策: 强化复杂多步骤任务的自主执行能力,能处理从信息搜集到分析处理再到结果交付的完整链路
  • 本地安全沙箱: 数据和文件在用户设备本地处理,不随意外流,适合企业和个人处理敏感办公数据
  • PC与移动端实时互通: 7x24小时AI助手陪伴,PC端与移动端数据、任务状态实时同步,无缝切换
  • Skill技能体系: 上线面向特定场景的专业技能,如高考志愿填报信息助手Skill,采取多源搜索加交叉验证机制,确保数据准确性

差异化优势: 百度最大的壁垒在于搜索生态。DuMate在查询和处理所有涉及外部事实的数据时默认进行实时搜索,模型记忆仅用于理解用户意图,不会基于模型记忆做出事实性判断——这一机制有效抑制了AI幻觉,在需要高精度信息的场景中具有显著优势。百度创始人李彦宏提出的DAA(Daily Active Agents,日活智能体数)概念,也为智能体时代的价值衡量提供了新标尺。


三、字节Trae Work:从编程工具到AI工作空间的进化

核心定位:让TRAE为你工作

Trae Work的前身是字节跳动自研的AI原生IDE——Trae。2026年3月,Trae推出"SOLO独立端",包含Code和Work两种模式。6月9日,SOLO正式升级更名为Trae Work,产品定位从独立开发者工具扩展为面向各类专业人士的AI工作空间。

关键数据: - 累计注册用户突破600万,月活跃用户达160万 - 2026年Q2以约1279万次月访问量位居国内AI办公智能体市场第二 - 国内版完全免费,无Token限制

核心能力:

Trae Work的独特之处在于它从编程工具基因出发,将"代码执行"能力扩展到更广泛的办公场景:

  • 双模式架构: IDE模式(基于VS Code同源架构,人机协同编码)+ Work模式(原SOLO模式,AI主导执行,人审核验收),覆盖从单行补全到全项目自动生成的完整链路
  • Builder+Coder双智能体: Builder负责"从0到1"项目搭建(自动生成PRD、架构、代码),Coder专注"从1到N"迭代(跨文件检索修改,支持10万代码文件处理)
  • 极致中文适配: 中文语义理解准确率达98%,可精准识别口语化需求与中文注释
  • MTC协议工具生态: 已接入1.1万个工具,涵盖浏览器自动化、系统工具、专业工具等
  • 内置多模态生成: 将Seedream 5.0(文生图)和Seedance(文生视频)直接内置,一句话即可生图、生视频
  • 连接器市场: 开放第三方平台接入,如通达信MCP作为首家金融数据提供商入驻

差异化优势: Trae Work的根基是"代码执行力"。相比从对话助手起家的产品,它在技术执行层面的深度更强——从IDE编码到Work模式的全场景办公,这条演进路径让它在处理需要代码逻辑的任务(数据分析脚本、自动化流程、前端页面生成)时具有天然优势。更名SOLO为Work,本质上是把已经发生的开发者使用场景正式写进产品定位,让同一套能力换了一个更大众的入口。


从对话到执行:智能体时代的范式转移

AI智能体从对话到执行

三款产品的涌现,共同指向了一个清晰的行业趋势:AI正在从"能聊天"走向"真干活"。

过去几年,AI产品的想象是——用户在聊天框里说出目标,AI调用工具替人完成工作。但2026年的现实揭示了一个更深层的认知:自然语言确实可以统一工作的入口,但自然语言无法统一工作的成果。一份报告最终仍要以文档交付,一份经营分析仍需要表格和图表,一个项目仍需要任务分工与版本管理。

正因如此,腾讯、百度、字节三家虽然切入点不同——腾讯从AI编程助手走向个人工作助手,百度从搜索生态走向通用智能体,字节从开发者IDE走向全场景工作空间——但最终都开始增加文档、设计、项目管理、协作和版本管理能力。


三者对比:不同基因,不同路径

维度 腾讯WorkBuddy 百度DuMate 字节Trae Work
起点 AI编程助手→个人工作助手 搜索生态→通用智能体 开发者IDE→编程Agent→工作空间
核心引擎 混元Hy3 + 多模型 文心大模型 + 搜索生态 豆包 + DeepSeek + 多模型
生态壁垒 微信/企业微信/腾讯文档 百度搜索/秒哒/伐谋/百科 飞书/MTC工具链/Seedream
技术特色 全流程自主执行 长程任务+主动决策+实时搜索防幻觉 代码执行力+Builder/Coder双智能体
用户规模 月访问2097万(第一) WAIC镇馆之宝 月访问约1279万(第二)
多端覆盖 iOS/Android/鸿蒙/PC/小程序 iOS/Android/PC 桌面端/网页端
数据安全 本地沙箱+信通院认证 本地安全沙箱 云端算力+本地Workspace
免费策略 签到领积分,Hy3限免 部分免费 国内版完全免费

结语:入口之争才刚刚开始

腾讯WorkBuddy负责人刘毅说得很坦诚:"上半年大家还是处于起步阶段,暂时跑赢半个身位而已。市场仍然处于非常早期阶段。"

这句话同样适用于整个赛道。2026年6月6000万次的月访问量,相较于中国数亿办公人群,渗透率仍然很低。三家巨头几乎在同一时间收拢产品线、统一入口——腾讯将QClaw团队并入WorkBuddy体系,阿里整合QoderWork/悟空/MuleRun推出"千问办公",字节推进飞书/TRAE/豆包多线并进——标志着行业从"各自试错"切换到"集中推进"。

WorkBuddy凭借腾讯生态和先发优势暂时领跑,DuMate以搜索护城河和SOTA级执行力站稳脚跟,Trae Work则以代码基因和免费策略快速追赶。三足鼎立的格局已经形成,但终局远未到来。

在智能体时代,真正的赢家不是技术最强的那个,而是最能理解"工作本质"的那个——把AI从"能说"变成"能干",从"给建议"变成"交成果"。这场关于人机协作入口的战争,才刚刚打响第一枪。

Sign up for Tencent orkbuddy to get 2000 points, which can be used for quite a few tasks. Download link: https://www.workbuddy.cn/

Baidu Dumate intelligent agent, also a good work companion tool: invitation code and download address: Baidu Dumate invitation code: 3V3JPSD

Bytedance TraeWork offers 4500 free points; click to download Bytedance TraeWork free 4500 points.

In the summer of 2026, China’s AI office track will experience a true turning point.

According to data from iResearch, the monthly visits of 17 mainstream desktop-based AI native office intelligent agents in China exceeded 60 million by June 2026 — whereas the figure was only 20 million in March. In just three months, user volume tripled. Office intelligent agents are evolving from "trial tools" to daily productivity instruments for workplace professionals.

In this battle for the entry point, three products — Tencent WorkBuddy, Baidu DuMate, and Bytedance Trae Work — represent three distinct technological paths and ecological strategies. They reflect how three internet giants answer the question: "How does AI reshape the workplace?"

I. Tencent WorkBuddy: From "One Sentence" to "Delivery" Across All Scenarios

Core positioning: A full-scene workplace AI intelligent agent desktop work platform

WorkBuddy was officially launched in March 2026 as Tencent's flagship AI office intelligent agent product. According to Tencent’s Q1 2026 financial report, based on daily active accounts, WorkBuddy has become China’s most popular efficiency AI intelligent agent service.

Key data: - June 2026 monthly visits reached 20.97 million, ranking first in China’s AI office intelligent agent market. - Surpassed the combined monthly visits of the second and third-ranked competitors. - On July 18, at WAIC 2026, an independent app was released, making it the first universal intelligent agent application to simultaneously launch on iOS, Android, and HarmonyOS.

Core capabilities:

WorkBuddy’s greatest feature is its "end-to-end autonomous execution capability" — it is not a chatbot but a digital employee that actively performs tasks for you. Faced with complex demands like "global population structure change visualization analysis," WorkBuddy autonomously plans steps, reads and analyzes local and online data, calls drawing skills, and completes layout and dashboard development to deliver professional results.

Flexible model switching: It is fully connected to mainstream models like Tencent Hunyuan Hy3, MiniMax, Kimi, DeepSeek, and GLM, and even supports API integration with private models. After the release of Hunyuan Hy3, over 60% of users opted for this model.

Autonomous decomposition and closure of complex tasks: It not only understands requirements but excels in executing them — autonomously planning steps, calling tools, and iterating until completion.

Local file autonomous management: With authorization, it directly reads and writes local folders, instantly performing bulk renaming, cross-table data extraction, and automatic file archiving.

Ecosystem connectivity: It is integrated with Tencent Documents, WeChat, and other Tencent services; it features a skill center, expert mode, automated scheduled tasks, and a knowledge base.

Multi-device collaboration: Supports PC, app, and WeChat mini-program platforms for remote task execution.

Differentiated advantage: Tencent's complete social and office ecosystem (WeChat, WeChat Work, Tencent Documents, Tencent Meeting) provides a natural connection advantage for WorkBuddy. The Hunyuan Hy3 model, released on July 6, achieves "compact size with flagship performance" with 295B total parameters and 21B active parameters, proving capable of handling 90% of agent tasks — a powerful engine for WorkBuddy.

II. Baidu DuMate: From "Idea" to "Result" with Super Executive Power

Core positioning: A true worker, your office buddy

Baidu DuMate released its desktop version in March 2026 and launched its mobile app on May 13 at the Create2026 Baidu AI Developer Conference. DuMate achieved SOTA levels in multiple international authoritative agent benchmarks and was selected as a "hall exhibit" at WAIC 2026, the only universal intelligent agent to receive that honor.

Core capabilities:

DuMate’s core selling point is "super executive power" — a single sentence triggers automated execution from idea to result. It integrates core capabilities from the Baidu ecosystem into callable built-in skills.

Deep integration with Baidu ecosystem: Incorporates Baidu AI Search, Miaoda, Famu, and Baidu Encyclopedia into built-in skills via a unified interface. Unlike agents that call third-party tools, Baidu's complete ecosystem from base model to vertical agent ensures smoother tool invocation and higher accuracy.

Long-range task execution and active decision-making: Strengthens autonomous execution of multi-step tasks, handling the full chain from information collection to analysis and delivery.

Local security sandbox: Data and files are processed locally on the user’s device, ensuring sensitive office data is not exposed to the outside.

Real-time synchronization between PC and mobile: A 7x24 hour AI assistant that seamlessly switches between PC and mobile with synchronized data and task status.

Skill system: Offers professional scene-specific skills, such as the College Entrance Exam Volunteer Information Assistant Skill, which uses multi-source search and cross-validation to ensure data accuracy.

Differentiated advantage: Baidu’s search ecosystem sets it apart. DuMate performs real-time searches for all external facts during queries and processing, using model memory only for intent understanding — a mechanism that effectively suppresses AI hallucinations and ensures high precision where factual reliability is critical. Baidu founder Robin Li’s concept of DAA (Daily Active Agents) also introduces a new metric for evaluating the value of intelligent agents.

III. Bytedance Trae Work: Evolution from Programming Tool to AI Workspace

Core positioning: Let TRAE work for you

Trae Work originated from Trae, a Bytedance-developed AI-native IDE. In March 2026, Trae introduced the "SOLO" independent end with Code and Work modes. On June 9, SOLO was officially upgraded and rebranded as Trae Work, expanding its scope from an isolated developer tool to an AI workspace for all professionals.

Key data: - Cumulative registered users exceed 6 million, with 1.6 million monthly active users. - In Q2 2026, it ranked second in China’s AI office intelligent agent market with approximately 12.79 million monthly visits. - The domestic version is fully free, with no token limits.

Core capabilities:

Trae Work’s uniqueness lies in extending the "code execution" capability of its programming tool roots into broader office scenarios.

Dual-mode architecture: IDE mode (based on VS Code for human-machine collaborative coding) and Work mode (formerly SOLO, AI-driven execution with human review), covering the full spectrum from single-line completion to full-project automation.

Builder+Coder duo: Builder handles "0 to 1" project setup (automatically generating PRD, architecture, and code), while Coder focuses on "1 to N" iteration (cross-file retrieval and modification, supporting up to 100,000 code files).

Extreme Chinese adaptation: Chinese semantic understanding accuracy reaches 98%, capable of precisely recognizing colloquial requirements and Chinese annotations.

MTC protocol tool ecosystem: Access to 11,000 tools, including browser automation, system tools, and professional instruments.

Built-in multimodal generation: Integrates Seedream 5.0 (text to image) and Seedance (text to video), enabling image and video creation with a single command.

Connector market: Opens up for third-party platform integration, such as Tongdaxin MCP becoming the first finance data provider to join.

Differentiated advantage: Trae Work is fundamentally built on "code execution power." Compared to products rooted in conversational assistants, it has deeper technical execution capabilities — from IDE coding to full-scenario Work mode — giving it a natural edge in tasks requiring coding logic (data analysis scripts, automation workflows, front-end page generation). The rebranding from SOLO to Work reflects the actual use cases of developers, opening the same powerful toolset to a wider audience.

From Conversation to Execution: The Paradigm Shift of the Intelligent Agent Era

AI智能体从对话到执行

The emergence of these three products reveals a clear industry trend: AI is transitioning from "being able to chat" to "truly doing the work."

For years, the vision for AI was that users would state their goals in a chat box, and the AI would call tools to complete the task. The reality of 2026 shows that natural language can unify the entry point, but not the output. A report must ultimately be delivered as a document, a business analysis requires tables and charts, and a project demands task assignment and version control.

This is why Tencent, Baidu, and Bytedance — despite different starting points (Tencent from AI programming to personal work assistance, Baidu from search to universal agents, Bytedance from IDE to full-scenario workspaces) — are all adding document, design, project management, collaboration, and versioning features.

Comparative Analysis: Distinct Genes, Divergent Paths

Concluding Remarks: The Battle for the Entry Point Has Just Begun

As Tencent WorkBuddy’s lead admits, "In the first half of the year, we are all just getting started; a slight lead is insignificant."

This sentiment applies to the entire field. The current monthly visit count of 60 million is low relative to China's hundreds of millions of office workers, indicating significant potential for growth. The recent consolidation of product lines by all three giants — Tencent merging QClaw into WorkBuddy, Alibaba combining QoderWork/Wukong/MuleRun into "Qianwen Office," and Bytedance advancing Feishu/TRAE/Doubao — marks a shift from independent experimentation to concentrated development.

WorkBuddy leads with its ecosystem and timing, DuMate solidifies its position with search superiority and execution excellence, and Trae Work rapidly catches up with its coding heritage and free-to-use model. While a tripod-like structure is forming, the final winner is far from decided.

In the age of intelligent agents, success depends not just on technical prowess, but on understanding the nature of work — turning AI from a conversationalist into an executor, from an advisor into a results-provider. The war for the human-machine collaboration interface has only just begun.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry