WorkBuddy makes office agents an operating layer

Desktop AI office agents already have mass attention in China, but attention is becoming the wrong scoreboard. According to woshipm, Analysys data showed that monthly visits to 17 domestic desktop AI office agents exceeded 60 million in June 2026. The same account cites China Internet Network Information Center data showing 602 million generative AI users as of December 2025, close to half of China's internet users.

That scale does not prove adoption.

The harder test is whether an agent becomes part of the operating layer for a task: the place where work starts, data is read, decisions are suggested, and payment can eventually sit. IDC research cited by woshipm found that only 18% of enterprises had truly put AI agents into core business processes, while more than 60% were still evaluating or piloting them. That gap explains the new fight. The prize is not the cleverest demo; it is control of the entry point and the routine around it.

Traffic is a weak test for an agent

Traffic makes office agents look mature before the work itself has changed. According to woshipm, Analysys data showed that in June 2026, monthly visits to 17 domestic desktop AI office agents exceeded 60 million. Another woshipm-cited report from the China Internet Network Information Center put China at 602 million generative AI users as of December 2025, close to half of all internet users. That is a huge attention base, but it is still only the top of the funnel.

WorkBuddy sits at the sharp end of that contradiction. In the AI office agent traffic ranking cited by woshipm, WorkBuddy had more than 20 million monthly active users, and its monthly active users exceeded the combined total of the second- and third-ranked products. The scale matters because office software has always been habit-bound. If a product can become the place where documents, meetings, approvals, and search begin, traffic can turn into control.

But visits are a weak test.

The harder question is whether the agent reaches the work process that a company cannot easily abandon. IDC research cited by woshipm showed that only 18% of enterprises had truly incorporated AI agents into core business processes, while more than 60% were still at the evaluation or pilot stage. That gap explains why the office-agent boom can look large from the outside and unfinished from the buyer's side.

Retention shows the same problem at user level. Woshipm cited data indicating that the 7-day retention rate for AI tool products is generally around 10%, and its author argues that AI office products still have not answered a basic user question: what exactly can I use this for? The author's useful frame is three thresholds from trial to reuse: workflow embedding, data integration, and habit migration. Without those, FOMO can bring users in, but it cannot make the agent an operating layer.

The platform fight is really an entry-point fight

The office-agent race is becoming a fight over the first screen a worker opens, not a beauty contest among model demos. According to woshipm, the first half of 2026 saw Tencent merge QClaw into WorkBuddy, ByteDance deeply integrate Feishu and Doubao, Alibaba fold QoderWork, Wukong, and MuleRun into Qianwen Office (千问办公), Baidu gather related products under Dazi, and Kingsoft Office launch AI-native products such as Lingxi Professional Edition and WPS Comate.

That pattern can be read as fear of missing out. If rivals are turning office suites into agent surfaces, nobody wants to leave a standalone assistant outside the daily work loop.

San Ge's sharper frame is platform consolidation. He argues that the platform-level entry-point war for AI office agents has begun among Tencent, Alibaba, and ByteDance. Alibaba's case shows the logic clearly: Qianwen Office 0.1.0 beta launched on July 27, with DingTalk's new CEO Chen Yusen leading it. QoderWork became the core, while Wukong and MuleRun were folded in, giving Qianwen Office a unified entry point and agent foundation.

ByteDance's restructuring points the same way. Woshipm says ByteDance announced on July 30 that Feishu would be split and integrated: the Feishu product team joined the Doubao product team, while Feishu's market, sales, and customer teams moved into Volcano Engine to form the Long Creativity Service Platform. Doubao Enterprise Edition, also called Doubao Office, is in beta testing, with Feishu's team directly involved.

OpenAI made a similar surface choice by merging the standalone Codex app into the ChatGPT desktop client on July 9, 2026. Tencent's July internal notice did the same organizational work around WorkBuddy. The agent has to live where demand can be captured.

How Chinese coverage defines agents becoming operating layers

Data access creates usefulness and lock-in at the same time

An office agent becomes useful only after it can see the work it is supposed to change. That is the practical meaning behind woshipm's cited IDC research: only 18% of enterprises had truly incorporated AI agents into core business processes, while more than 60% were still evaluating or piloting them. The gap is not mainly about curiosity. It is about access.

woshipm's author describes three thresholds between a trial and repeat use: workflow embedding, data integration, and habit migration. For office agents, that means the agent must sit close to documents, calendars, enterprise chat, code tools, and the daily client where employees already start work. A chatbot that waits outside those systems can draft a paragraph. An embedded agent can notice the meeting, find the contract, update the task, and call the right tool without making the user rebuild context each time.

That usefulness is also the lock-in.

Alibaba's Qianwen Office (千问办公) shows the pattern. According to woshipm, Alibaba integrated QoderWork, Wukong, and MuleRun into the Qianwen Office Agent, with QoderWork as the core and Wukong and MuleRun adding capabilities. The product is described as a unified entry point and agent foundation. Once the entry point owns the agent foundation, the office client is no longer just where files are opened. It becomes the place where permissions, task memory, and tool calls accumulate.

ByteDance is moving in the same direction from another base. woshipm reported that the Feishu product team was integrated into the Doubao product team, while Feishu's market, sales, and customer teams were integrated into Volcano Engine to form the Long Creativity Service Platform. Doubao Enterprise Edition, also described as Doubao Office, is being developed with direct participation from the Feishu team. The signal is clear: enterprise distribution and office habits are being folded into the agent product itself.

The same consolidation is visible outside China. OpenAI merged the standalone Codex app into the ChatGPT desktop client on July 9, 2026. Tencent issued an internal notice in July that the QClaw team and business would be merged into the department where WorkBuddy is located. woshipm's author San Ge argues that the platform-level entry point war for AI office agents has begun among Tencent, Alibaba, and ByteDance. The prize is not a single feature. It is the working layer employees would have to leave behind.

The agent layer may become the ad layer

If the entry point moves from a search box or app icon to an agent, the old traffic bargain starts to wobble. According to ifanr, Elon Musk recently agreed with a Cloudflare report that most internet traffic in the near future may no longer come from humans, but from AI and agents operating in the background. The author's sharper point is that traffic can keep growing while becoming traffic without eyeballs. That matters because web advertising was built around human attention.

Agents can fetch, compare, summarize, and transact while showing the user fewer pages. That makes control of the entry point less secure than it looks.

The ad business will not simply disappear. ifanr argues that AI may become advertising's new host, even a near-ideal salesperson. Google will place sponsor advertisements, including products or information, directly inside AI-generated answers in AI Overviews. Perplexity has tried a sponsored questions format that steers users toward advertiser-funded follow-up questions. The incentive is obvious: previous data showed follow-up questions accounted for 40% of all Perplexity request traffic.

The same pressure appears outside chat. BMW faced controversy after overseas owners saw an advertisement for Sony and Marvel's "Spider-Man: Brand New Day" on the in-car central control screen after an in-car system update. The ad first appeared as text in a banner, could be dismissed, and played an animated video when opened. Overseas BMW models produced after July 2020 and running the seventh-generation or later operating system may show it.

That backlash was about broken expectations. BMW had repeatedly promised not to use dashboards and central control screens for marketing, and Stephan Durach had described the central control screen as private space rather than inventory to sell.

Windows shows the softer version of the same move. Microsoft added "recommended apps," described as sponsor advertisements, to the Start menu, and has inserted MSN feeds, game promotions, Bing advertisements, and other unrelated content into the lock screen. Roku's 2024 patent for "HDMI customized ad insertion" goes further in theory: it monitors HDMI signals and can insert ads when it detects a pause event, though Roku says it is only a technical reserve with no current launch plan.

So the agent layer is not only a gatekeeper for work. It can also become the place where ads survive after webpages lose attention.

When an AI agent is more than a demo

• Use an office agent when the product is becoming the user's daily entry point, not just a separate chatbot. Watch the platform integrations: Tencent merged QClaw into WorkBuddy, ByteDance integrated Feishu with Doubao, and Alibaba consolidated QoderWork, Wukong, and MuleRun into Qianwen Office. These moves matter because the Chinese coverage frames the contest as a platform-level entry point war, not a model leaderboard.
• Use an agent for enterprise work only if it can cross from trial use into actual business processes. Treat pilots cautiously. IDC research cited in the sources says only 18% of enterprises had truly incorporated AI agents into core business processes, while more than 60% were still at the evaluation or pilot stage. The practical test is whether the agent is embedded in workflows, integrated with data, and able to change habits.
• Use a cooking agent when the vendor controls appliances, recipes, usage data, and the physical kitchen workflow. Robam's Shishen AI cooking agent is presented as a kitchen "brain" and "operating system" that can understand needs, sense kitchen status, plan tasks, dispatch appliances and robots, adjust processes in real time, and learn through use. The case is strongest where the AI is tied to Robam's 827 AI-equipped product models, cooking-curve database covering more than 5000 dishes, and Shishen app and mini program with more than 7.5 million registered users.
• Do not assume user curiosity equals durable adoption. The sources describe heavy interest but weak proof of repeat use: Analysys data showed that in June 2026, monthly visits to 17 domestic desktop AI office agents exceeded 60 million, yet some data indicates that the 7-day retention rate for AI tool products is generally around 10%. The user question remains: "What exactly can I use AI office products to do?"
• Be cautious when an agent or device platform becomes an advertising surface. The sources connect agents and operating layers to new ad inventory: Google will place sponsor advertisements directly inside AI-generated answers in AI Overviews, Perplexity launched sponsored questions, and BMW owners saw a "Spider-Man: Brand New Day" advertisement on the in-car central control screen. If the workflow entry point is controlled by a platform, monetization may arrive inside the task itself.

Robam shows the promise and trap of vertical agents

Robam Appliances makes the cleanest case for a vertical agent because cooking gives Shishen a narrower job than a general office assistant. According to 36kr, Robam has focused on cooking for 47 years, serves 50 million global household users, and has a cooking-curve database covering more than 5000 dishes. Its Shishen app and mini program have more than 7.5 million registered users, with active users up 70% year-on-year.

That is the agent thesis in concrete form.

Shishen is presented as a kitchen brain and operating system: it can understand user needs, sense real kitchen status, plan cooking tasks, dispatch appliances and robots, adjust processes, and learn through use, according to 36kr. This matters because the hard part for agents is not answering once. Woshipm cites IDC research showing only 18% of enterprises had truly incorporated AI agents into core business processes, while more than 60% remained at evaluation or pilot stage. Woshipm's author argues reuse requires workflow embedding, data integration, and habit migration.

Robam has all three, but inside its own kitchen. The company says transformation is not just adding Wi-Fi or a smart label to old products. It has 827 product models equipped with AI functions, 6612 valid patents accumulated over nearly six years, and AI digital kitchen appliance annual sales of 2.35 billion yuan in 2025, up 36% year-on-year. Its i series increased 65% year-on-year.

The same depth creates the trap. Shishen becomes more useful as it touches recipes, sensors, appliances, robots, classrooms, and family routines. At AWE 2026, Robam unveiled AI cooking glasses and an AI seasoning robot; the glasses have been used in elementary school labor education food education courses in Hangzhou, Zhejiang Province. Its food education model is planned for no fewer than 1000 practice bases nationwide, and its Warm Kitchen Plan has helped nearly 7000 families in difficulty.

That is stronger than a demo. It is also harder to separate from Robam's own devices, data, sales channel, and teaching system. A vertical agent may beat a general one at a real task, but the price is that the operating layer can become a vendor's kitchen.

Treat the next office-agent pitch as a control question, not a feature question. If Doubao Enterprise Edition arrives from a Feishu-participating team, as woshipm reports, ask which Feishu data it can read, which actions it can complete, and which admin boundary it crosses. If WorkBuddy absorbs QClaw's team and business inside Tencent, ask what becomes default inside the desktop flow.

The entry point is the product.

Watch reorganizations as closely as demos. Feishu's market, sales, and customer teams moved into Volcano Engine to form the Long Creativity Service Platform, while Doubao's new product team sits under the head of Doubao. OpenAI made the same kind of move when it folded the standalone Codex app into the ChatGPT desktop client on July 9, 2026. Your test is simple: does the agent remove repeated steps in your real files, or does it move your work into someone else's operating layer?

For readers outside China

• Availability: The source material focuses on China-market products and Chinese tech-platform moves. It does not disclose availability outside China for WorkBuddy, Qianwen Office, Doubao Enterprise Edition, Feishu integrations, or Robam's Shishen AI cooking agent. Qianwen Office 0.1.0 beta launched on July 27, and Doubao Enterprise Edition is in beta testing and is expected to launch soon.
• Pricing: Pricing for the office agents and Shishen AI cooking agent is not disclosed in sources. The sources do give business-scale figures for Robam: in 2025, Robam Appliances' AI digital kitchen appliance annual sales reached 2.35 billion yuan, up 36% year-on-year, and its i series increased 65% year-on-year. Robam and China Academy of Art also jointly created a global cooking art center with an investment of 200 million yuan.
• Closest Western equivalents: Microsoft Copilot-style office agents embedded into productivity software and operating-system surfaces; ChatGPT desktop client-style consolidation of standalone agent tools into a main assistant entry point; Smart-home and connected-kitchen ecosystems that combine recipes, sensors, appliances, and automation; AI search and answer products such as Google AI Overviews and Perplexity, especially where sponsored answers or follow-up questions become part of the interaction
• Data residency: The sources do not disclose data-residency, hosting, cross-border data-transfer, or enterprise compliance details for the office agents or Robam's Shishen AI cooking agent. They do emphasize that the adoption barrier is data integration: Chinese coverage argues that AI office users must cross workflow embedding, data integration, and habit migration before trial use turns into repeat use.

Sources

• ifanr AI 即将接管流量，而广告正在疯狂「末日反扑」 https://ifanr.com/1676208
• woshipm AI办公也需要"牛来" https://woshipm.com/share/6450515.html
• woshipm AI 办公智能体，上演三国杀，WorkBuddy 遭字节阿里围攻 https://woshipm.com/share/6438261.html
• 36kr 科氪 | 当AI走进厨房，老板电器加速迈向AI烹饪科技企业 https://36kr.com/p/3945943299783812