Doubao Work vs WorkBuddy vs Qianwen Office: Comparing Three Desktop AI Agents

① Signal: WorkBuddy reaches 1 million daily active users, forcing ByteDance’s 30‑day integration

According to a August report by Jiemian News, WorkBuddy’s monthly active users have risen to the 20 million level, with daily active users stable at one million. In the second‑quarter iResearch report, WorkBuddy ranked first with 20.97 million PC‑side monthly visits, surpassing the combined total of ByteDance’s TRAE and Alibaba’s QoderWork.

The growth speed is notable: WorkBuddy entered public beta in March and climbed from about 8.85 million visits in its launch month to 20.97 million by June. Tencent promoted it aggressively, with internal rumors describing it as the third strategic‑level product after QQ and WeChat. This scale demonstrates that a general‑purpose agent can truly embed itself in real work flows.

The immediate consequence was a lightning‑fast organizational response from ByteDance. On August 25, Doubao Work was officially released, following a month of intense internal integration.

② What Doubao Work does: autonomous task decomposition, virtual desktop operation, multi‑Agent teaming

Content generation and editing : can produce documents, spreadsheets, PPTs, web pages, images and videos, supporting selective region editing without regenerating the whole file.

Computer and browser manipulation : after authorization, it can handle cross‑software data processing, fill forms, compare information, and continue long tasks on a cloud‑desktop even when the local machine is shut down.

Remote control via mobile : users can issue commands and check progress from their phones while away.

Multi‑Agent teaming : offers over 100 “work partners” such as UI designers, HR assistants and research analysts that can collaborate in small squads.

Doubao Work’s Windows “virtual desktop” uses GUI simulation to mimic human interaction with local software; tasks run in an isolated environment, do not interfere with the foreground, and can be manually taken over at any time. This moves it beyond the traditional “text‑answer” AI assistants.

③ ByteDance’s three‑step integration: Feishu for context, TRAE for execution, Coze for ecosystem

From July 30 to August 25, ByteDance completed organizational restructuring, capability integration, and brand independence in under 30 days.

Step 1 – Feishu adds context : On July 30, the Feishu product team merged into Doubao, with Feishu leader Xie Xin reporting to Doubao leader Zhao Qi. Feishu’s chat records, documents, meeting minutes, schedules and approvals provide the crucial “context” that agents lack.

Step 2 – TRAE adds execution : On August 24, the TRAE and Coze (formerly Koudi) teams were fully incorporated. TRAE’s IDE/CLI remains as a programming product line, supplying “end‑to‑end” engineering capability.

Step 3 – Coze adds ecosystem : Also on August 24, Coze brought an agent development platform and skill ecosystem, enabling external developers to contribute capabilities.

Between August 17 and 21, Doubao released a week of intensive updates: mobile remote‑control, Windows virtual desktop, side workbench, skill store, connectors, and over 200 skills were made available, culminating in the official launch on August 25.

④ Three‑way comparison: WorkBuddy relies on ecosystem breadth, Qianwen Office on DingTalk depth, Doubao Work on Feishu context

WorkBuddy – ecosystem breadth : Natively integrates Tencent Docs, Enterprise WeChat, Tencent Meeting, IMA, Tencent LeXiang, and can invoke WeChat and Enterprise WeChat, placing it directly in everyday social and work scenarios. PC‑side monthly visits grew from 8.85 million to 20.97 million after the March launch, showing clear traffic advantage.

Qianwen Office – DingTalk depth : Publicly tested on August 3, it consolidates the former QoderWork, MuleRun and Wukong agents. It enjoys a “bidirectional embedding” with DingTalk, accessing messages, group chats, schedules, to‑dos, documents, knowledge base, attendance and approvals, and can push results back to DingTalk. For over 20 million enterprise organizations, this native integration creates strong stickiness. It runs the Qwen 3.8 model and open‑sourced the context infrastructure project MyContext on August 18.

Doubao Work – Feishu context, serving both C‑side and B‑side : Its core differentiator is “understanding goals where work happens”. After logging in with a Feishu account, the agent can directly call real‑time context within Feishu’s authorized scope, eliminating the need for users to gather scattered data. Content co‑created with Doubao Work is continuously persisted back to Feishu as enterprise knowledge.

The three strategies essentially pick different entry points to the same endgame: Tencent targets “where people are”, Alibaba targets “where organizational processes reside”, and ByteDance targets “where knowledge is stored”.

⑤ Decisive point: personal entry is easy to capture, organizational context is the long‑term moat

36Kr highlighted a key insight: “Capability decides whether an agent can work; context decides whether it can work well.” As functional lists converge, the real friction moves from execution to preparation – gathering chat logs, latest version of plans, background information, and re‑explaining them to the AI.

Deloitte’s 2026 global survey shows AI tool coverage for employees rose from under 40 % to 60 %, yet only 34 % of companies use AI to deeply transform products, core processes or business models, while 37 % remain at superficial usage. The gap lies in context.

Feishu’s advantage is that it naturally unifies chat, documents, meetings, knowledge base, multi‑dimensional tables, projects and approvals under a single identity and permission system. Agents therefore do not need to move data across systems; they can find information, understand relationships, and act directly where work occurs. This explains why ByteDance absorbed Feishu – not to create another standalone app, but to weld the agent into the organization’s knowledge flow.

The narrative shifts: the first half of the competition compares who has the more capable agent (models, tool calls, long‑task handling); the second half compares who can truly embed the agent into organizational knowledge and workflows. Capability will quickly converge; context and workflow depth become the moat.

⑥ Verdict: the battle ultimately measures delivery of verifiable objective value

According to iResearch, the total monthly visits of 17 domestic desktop AI office agents exceeded 60 million in June, up from 20 million in March – a three‑fold increase in three months. With total desktop monthly active users around 30 million, traditional office software scale suggests roughly 20 times more room for growth.

Enterprise concerns remain concrete: ROI, data security, and accuracy on complex tasks. The era where AI only provides text answers and users must implement the output is ending; agents that can generate editable documents, operate real systems, and persist completed flows as reusable Skills will be treated as genuine productivity tools.

All three giants are fighting for the definition of “production tool”. WorkBuddy, with six months of head start, leverages ecosystem and social entry to drive volume; ByteDance closed its gaps in 30 days through rapid integration; Alibaba leans on DingTalk’s organizational lock‑in to gradually close the gap. Short‑term, the winner may be the one with the most complete feature set and strongest marketing; long‑term, the winner will be the agent that becomes the “colleague who never needs re‑briefing” within an organization.

In essence, the office agent’s endgame is not another chat window but a digital colleague that can handle real work flows, retain organizational context, and deliver verifiable outcomes – and that battle has only just begun.