Tencent WorkBuddy, Alibaba's QwenWork (千问办公) and Alibaba's QoderWake are the three most developed products in China for putting an AI agent into the group chats colleagues already use: WeCom, DingTalk and Feishu. Unlike Meta Muse or OpenAI dots, they had to answer on day one whether colleagues can use the agent. Their answers differ a good deal.
This comparison uses each vendor's site, documentation and changelog as of 30 September 2026. The Chinese original was written first. For the four products launched in the US, see personal agents in 2026, compared.
Three ways into the group chat: IM remote-controlling a desktop agent, three layers of agent inside DingTalk, and one bot the whole group mentions that routes to several Wakers.
What are the three products?
Tencent WorkBuddy comes from Tencent Cloud's CodeBuddy team. It launched in China on 9 March 2026 and internationally on 28 May, as "AI Agent for Everyday Office Work". It is a desktop app that runs on your computer and can "read authorised folders on your computer and batch-process them". It is compatible with OpenClaw skills, and version 5.6.2 added computer use. You can reach it from WeChat, WeCom, QQ, Feishu, DingTalk, Slack, Telegram and Discord. Each of these is a remote control for the WorkBuddy running on your computer.
Alibaba's QwenWork merged three products: QoderWork, a desktop agent; Wukong, DingTalk's enterprise agent platform; and MuleRun. The merger was led by DingTalk's CEO, Chen Yusen. The public beta opened on 3 August alongside Qwen3.8, and the product was upgraded at the Yunqi conference on 22 September. Alibaba calls it the first product to support desktop, cloud and enterprise-collaboration agents at once. The desktop layer handles local files and apps, the cloud layer runs research and long tasks, and the DingTalk layer handles group chats, approvals and to-dos.
Alibaba's QoderWake is part of the Qoder brand. Its public beta opened on 26 May and version 1.0 shipped on 3 September, under the line "Autonomous AI Employees, On the Job". It installs a local service with a web console and a CLI, where you build a team of digital employees called Wakers. It connects to DingTalk, Feishu and WeCom, and to WeChat for personal accounts. It can also be deployed to a cloud desktop or a Linux server so the Wakers do not depend on your computer.
How do they compare?
How do colleagues actually use the agent?
WorkBuddy uses a visibility scope. When you create the WeCom bot, the docs ask you to set which employees, departments or tags can use it. Those colleagues can then talk to your bot, and behind the bot is the WorkBuddy on your computer. It is the lightest of the three answers: no new workspace, no new accounts, one setting that turns "my agent" into "my department's agent". The cost is that the permission stops at who can use it. The docs do not say whether different colleagues can do different things.
QwenWork uses an organisational identity. A digital employee has a name, a department and a responsible person like any member of staff, with written duties, authorisation limits and a traceable log. A task can come up in a company group chat, be carried on by the cloud agent, and then call on files and apps on an employee's computer. It is the most enterprise-shaped answer, and it assumes your company already runs on DingTalk.
QoderWake uses one bot with a team behind it. Everyone in the group mentions the same bot, which works out what each person wants and routes the task to the right Waker. The docs are explicit that names like project coordinator, developer or QA are for the division of work behind the scenes and "do not become multiple bots in the group". A group cannot use the bot just because the bot is in it. An admin has to approve the chat's request, or someone has to pair it with a code that is usually valid for ten minutes. The docs warn not to post pairing codes in unrelated groups or keep them.
What is wake-up mode?
Wake-up mode (觉醒模式) gives QoderWake its name, and it marks the line between QoderWake and an ordinary group bot. It is a separate switch for each chat. When it is off, the chat is a plain messaging connection. When it is on, the chat gets the full @Waker mode: tasks are claimed and routed, a record is kept, and topics and memory build up over time.
So the same bot can be two different things in two groups. In a social chat it only answers. In a project group it is a team of digital employees that claims tasks and remembers context. Once a chat has two or more Wakers, the docs recommend running a duty-conflict check, because overlapping responsibilities can send similar tasks to different Wakers.
Does the computer have to stay on?
For WorkBuddy and QoderWake, by default, yes. WorkBuddy's messaging connections remote-control the desktop app, which is why it has a setting to stop the computer sleeping. Its mobile app has a cloud mode that needs no computer, and its enterprise digital employees can be deployed in the cloud. QoderWake's docs say plainly that the device must be online when a task needs local files, and suggest a cloud desktop or server if you do not want to depend on your own machine. QwenWork's cloud layer never depended on your computer, and it calls the desktop layer only when it needs local files.
Running locally means the files and logins are already there, and no credentials go to the cloud. The downside is that a sleeping machine means a sleeping agent. The US-launched products, Muse, dots and Grok Bot, take the opposite default and run on the vendor's cloud computer.
What do all three leave open?
All three let colleagues use the agent. None of them spells out who approves when a colleague's request reaches an irreversible step. A visibility scope decides who can use the agent, an organisational identity decides what the agent may do, and a pairing code decides which chat may use it. But suppose a request comes from one colleague and runs on another's computer. Should a push, a delete, a payment or a credential read be confirmed by the person who asked, the person whose machine it is, or a rule? The documentation assumes the person asking and the owner are the same.
muti is built around that question. It is not a personal agent. It gives the coding agent already running on your computer, such as Claude Code or Codex, a room where the people you invite send it work. Each person has their own permission: read-only, edit or a standing grant. Four kinds of action always stop for the owner, whoever asked: pushing to a shared branch, deleting, spending money and touching credentials. The design of muti's Feishu connector, which is in development, is at a Feishu group that can @ the Claude Code on your machine. The idea behind it is at what is an agent room.
Which one fits which team?
Choose WorkBuddy if your team is on WeCom or WeChat and you want colleagues using the agent on your computer as quickly as possible. It covers the most messaging apps, takes the least setup, and the international Pro plan is $10 a month.
Choose QwenWork if the whole company is on DingTalk and you want AI employees with a formal identity, approvals and audit. Its three layers are the most complete, and tasks raised in a group chat can run on in the cloud.
Choose QoderWake if you want a team of digital employees with clear roles in one group, and strict control over which chats can use them. Wake-up mode, pairing codes and duty-conflict checks make it the most thorough of the three about groups.
Frequently asked questions
What is the difference between WorkBuddy and QoderWake?
Both run on your own computer by default. WorkBuddy is one agent with many messaging remote controls, and colleagues use it through a WeCom visibility scope. QoderWake hides a team of Wakers with distinct roles behind one bot that the group mentions. Each chat is enabled by an admin or a pairing code, and wake-up mode decides whether it is a plain connection or the full digital-employee mode.
How is QwenWork related to QoderWork?
QoderWork is the desktop agent Alibaba launched in January 2026. In July it merged with DingTalk's Wukong and MuleRun into QwenWork, where it became the desktop layer. In China, QwenWork is the product. Qoder's international plans still list QoderWork beside QoderWake.
Do they work with Feishu (Lark)?
WorkBuddy and QoderWake both have official Feishu connections. QwenWork centres on DingTalk, and its international edition connects Slack and Notion. Other messaging apps for QwenWork appear only in press reports, not in its own documentation.
If a colleague uses my agent, does it use my accounts and files?
Yes. When any of the three runs on your computer, it uses your files and your logins, and your machine executes the colleague's request. So the question is not whether colleagues can use it, but which actions must go through you first. Before you open it up, check that the high-risk rules also apply to requests from other people.
Sources
- WorkBuddy: workbuddy.ai · WeCom guide · TechNode
- QwenWork: qwenwork.ai · pricing · TMTPost
- QoderWake: qoder.com/qoderwake · overview · IM configuration · changelog
- SCMP on Alibaba's digital employees: scmp.com
If the agent worth sharing is already on your computer, let colleagues use it and keep the approval yours. Give it a room.
公開日