Author | Bishan

Source | Bowang Finance

On the afternoon of August 14, at Baidu's AI Day open house event, the Baidu Wenku and Cloud Storage team gave their general-purpose AI agent GenFlow, which boasts over 100 million monthly active users, a down-to-earth Chinese name—"Kuku AI"—and go with the flow (shùnshì, meaning " go with the flow 而为" or "seizing the opportunity") launched an independent client and enterprise version. Moving the clock forward, on August 3, Alibaba's enterprise-level Agent "Qianwen Office" entered public beta, and on the same day, they released their new flagship model Qwen3.8; on July 30, ByteDance issued an internal memo merging the Feishu product team into Doubao; on July 20, Tencent reassigned the QClaw product center to Cloud Product Division VI, where WorkBuddy resides.

In less than a month, four major internet companies have accomplished the same task: consolidating their internal AI product lines that have been racing independently for years and focusing their efforts on a unified direction—AI office solutions. For the first time, the question that has been repeatedly asked over the past few years—"Where is the next-generation gateway?"—has a highly consistent answer.

01

The End of the Dialog Box Era After Lobster

To understand why big tech companies are collectively shifting their focus now, one must first recognize the changes that occurred on ordinary users' computer desktops in the first half of this year.

OpenClaw, which ignited the tech community at the beginning of the year, was originally just a "weekend project" written by Austrian retired programmer Peter Steinberger in November last year. This open-source framework, which aimed to give large models access to local operating system permissions, went viral after its release in January 2026: its GitHub stars surged to around 260,000 within a few months, surpassing established projects like React and Linux; because its icon is a red lobster, Chinese netizens simply called the process of deploying and using it "raising lobsters." In early March, Tencent set up a stall outside its Shenzhen headquarters to help people install it for free, with nearly a thousand people lining up; in Ma Huateng's own words, he never expected it to become this popular.

Beyond its popularity, the lobster made a deeper impact on the industry by demonstrating a new possibility to ordinary users. In the past, AI stayed within dialog boxes—you asked a question, and it answered; however, desktop Agents like OpenClaw can directly take over browsers, local files, and terminal software, breaking down steps, invoking tools, and delivering results on their own. AI has transformed from "answering questions" to "executing tasks," and for office workers, the computer desktop has become the true workplace for large models.

If the lobster proved the technical feasibility of Agents, the explosion in demand is even more evident in the data. Liu Liehong, director of the National Data Bureau, revealed at the China Development Forum 2026 in March that China's daily Token calls were only 100 billion at the beginning of 2024, surging to 100 trillion by the end of 2025, and breaking through 140 trillion in March this year—a more than thousandfold increase in two years. Google's curve announced at I/O 2026 in May was equally steep—monthly Token processing surged from 480 trillion in the same period last year to over 3,200 trillion, a sevenfold year-on-year increase.

What's even more telling than call volume is the flow of money. Anthropic's programming Agent product Claude Code, publicly released in May 2025, achieved an annualized revenue of $1 billion in six months and doubled to $2.5 billion by February this year; Anthropic's own annualized revenue grew from $9 billion at the end of 2025 to $30 billion in April this year, with SemiAnalysis's May estimate reaching $44 billion. Currently, over 1,000 enterprise clients spend more than $1 million annually on Anthropic—these budgets are no longer aimed at chatbots. What businesses want are "digital employees" who can do the job.

02

Why Office Solutions: The First Economically Viable Scenario

While technological maturity is a prerequisite, what truly convinced big tech companies to consolidate their efforts is the economic viability of the office scene itself.

Comparing the various landing directions for large models, office solutions are almost the most certain. Tasks like document processing, data analysis, and code generation occur frequently, with delivery standards that are almost unforgivingly clear—a report, a spreadsheet, a PPT slide: right or wrong is immediately apparent; in contrast, vertical scenarios like industry have low fault tolerance and lengthy processes, making it difficult for Agents to deliver standardized results in the short term. Gartner previously predicted that by 2028, at least 15% of daily work will be autonomously completed by AI agents, and one-third of enterprise software applications will have built-in agent capabilities.

Market size calculations also point to this conclusion. The "2026 China AI Office Agent Industry Development White Paper" released by iiMedia Research provides a set of figures: China's AI agent market size reached 80.4 billion yuan in 2025, a 123.2% year-on-year increase, and is expected to approach 700 billion yuan by 2030.

Beyond the size of the cake, changes in how money is collected are also worth examining. The previous generation of office software charged subscription fees per seat, essentially selling accounts; in the Agent era, what businesses pay for is shifting toward Token consumption and task completion volume. Two overseas attempts provide references: Microsoft sells Copilot as a high-priced AI seat separately, but penetration has been lackluster; Google chose to bundle Gemini into Workspace packages; Anthropic and OpenAI, relying on API call billing, have achieved revenue growth rarely seen in the software industry's history. Pricing power is far from settled, and whoever first occupies the desktop gateway has the opportunity to help set the rules.

03

Four Tables, Four Approaches

While all four companies are eyeing the office solutions cake, the cards they hold differ, determining their respective routes.

Tencent's strength lies in "people"—or more precisely, the connections between people. WorkBuddy's technical foundation comes from Tencent Cloud's AI programming tool CodeBuddy, launched in 2023: after Claude Cowork's release in January this year, CodeBuddy's head, Wang Shengjie, and a few colleagues created the initial version over a weekend, first testing it internally with over 2,000 Tencent employees before its official public beta on March 9. According to Analysys's "2026 Q2 China Office Agent Platform Market Insight Report" released on July 20, WorkBuddy's PC-end monthly visits reached 20.97 million in June, ranking first among 17 mainstream desktop office agents, surpassing the combined total of Byte's TRAE (second place) and Alibaba's QoderWork (third place); Tencent's Q1 earnings report also disclosed active user retention exceeding 60% and paid user retention exceeding 80%. Strategically, WorkBuddy is not deeply tied to WeChat Work but is instead an independent desktop application, supporting remote activation from WeChat, WeCom, QQ, and even Feishu and DingTalk, spreading its gateway across all mainstream IM platforms. According to Jiemian News, this project has received green lights internally at Tencent, with Ma Huateng rarely missing product meetings; in the Q1 report, he directly wrote: "We believe WorkBuddy is currently China's most widely used efficiency AI agent service."

Unlike Tencent, Alibaba's strength lies in "organization." The public beta of Qianwen Office on August 3 integrates three products: QoderWork, Wukong, and MuleRun, led by Chen Yusen, a 1992-born technical manager who took over as DingTalk CEO in June. The integration background was not easy: public reports have reconstructed Alibaba's internal racing dilemma—over a dozen Agent products scattered across different teams, with overlapping positioning to the point where employees "basically only used QoderWork and rarely opened Wukong"; in early June, a 75,000-word resignation essay titled "Inside DingTalk" laid bare the contraction process of DingTalk's flagship AI project, followed by Chen Hang's resignation and Chen Yusen taking over, with integration swiftly advancing. The final Qianwen Office does not compete head-on with personal desktops and WorkBuddy but focuses on enterprise organizational scenarios: officially defined as the industry's first product to simultaneously support desktop agents, cloud agents, and enterprise collaboration agents, it has initially integrated with DingTalk IM and aims next to connect with enterprises' real databases and workflows—precisely what DingTalk has accumulated over a decade: organizational structures, approval processes, and permission systems.

ByteDance's approach bears its distinct DNA: using C-end growth logic to transform the B-end. In its July 30 internal memo, the Feishu product team was fully merged into Doubao, forming a new Doubao product team led by Doubao head Zhao Qi, with Feishu head Xie Xin reporting to him; Feishu's sales and marketing teams merged with Volcano Engine into a "Creativity Service Platform," uniformly handling MaaS and SaaS commercialization. Zhao Qi had previously overseen Douyin's growth and the Pangle advertising platform before taking over Doubao in September 2025. ByteDance's calculations are clear: according to Xin Huanghe·Dayu Finance, by June, Doubao's large model had exceeded 180 trillion daily Token calls, and based on July's average consumption, ByteDance's large model business ARR reached $4 billion; entrusting Feishu's accumulated enterprise organizational context to Doubao for scheduling and merging the sales system into Volcano Engine allows "office collaboration + large model + cloud services" to be packaged as a one-stop solution for enterprise clients. Currently, the Doubao enterprise version, deeply developed with Feishu's team, has begun internal testing among some Feishu clients.

Baidu takes a different path from the other three—cutting in from "data." Baidu Cloud Drive has over 1 billion registered users with a total user storage space exceeding 100 billion GB, and Baidu Wenku has accumulated 1.8 billion professional documents—files generated during work processes are precisely the raw materials Agents need most to get the job done. On August 5, the National Industrial Information Security Development Research Center, a subsidiary of the Ministry of Industry and Information Technology, released the "Office Agent Workflow Evaluation Report," assessing the Office workflow capabilities of five mainstream products on the market, with Baidu Wenku ranking first and alone in the top tier. In terms of user scale, after GenFlow's monthly active users surpassed 100 million in April this year, the team discovered that office solutions were its highest-frequency scenario; according to media reports, they assigned about 30 people to develop an independent office client in just over 20 days, officially announcing "Kuku AI" on August 14—currently, its AI office MAU exceeds 25 million, ranking first in the general-purpose AI office track. Beyond Kuku AI, the no-code platform Miaoda had served over 35 million users and generated 3.5 million applications by July this year; the desktop-level general-purpose Agent "Baidu Dazi" was selected as one of the top ten "museum pieces" at the World Artificial Intelligence Conference in July, the only general-purpose agent product to make the list.

04

After Model Convergence, Context Becomes the Key

Viewing the actions of all four companies together, an industry consensus is emerging: large models themselves are converging in capability, and model performance alone no longer determines product success.

Tencent's open-source evaluation benchmark WorkBuddy Bench, released in July, provides an observation window for this judgment. Developed jointly by teams including Tencent Youtu Lab, the benchmark contains 260 questions across four subsets—code, frontend, office, and security—all Reverse rewriting (nìxiàng gǎixiě, meaning "reverse-engineered") from real work scenarios, with the office subset requiring Agents to complete full workflows on mixed-format files like xlsx, csv, pdf, and docx. From the publicly released cross-model evaluation results, the score gaps among several leading models in office tasks are already very small—when everyone can score around 80, the exam itself is no longer the differentiator.

When model scores cannot widen the gap (lā kāi chājù, meaning "create differentiation"), the true differentiator shifts to context. For an Agent to truly "get the job done," a clever large model alone is far from enough: it must understand enterprise knowledge bases and historical files, connect to WeChat, email, calendars, and databases, invoke various tools within permission boundaries, and meet Mandatory requirements of enterprises (qǐyè de yìngxìng yāoqiú, meaning "enterprises' strict requirements") for data security and auditing. Where does this context come from? Tencent's accumulate (jīlěi, meaning "accumulation") comes from communication relationships and collaboration networks; Alibaba holds organizational and process data; ByteDance bets on Doubao's gateway combined with Feishu's collaboration scenarios; Baidu holds the file assets accumulated in Cloud Drive and Wenku. With different data types in hand, their routes naturally diverge.

Moreover, the value of data lies not just in its inventory (cúnliàng, meaning "stock") but in its ability to roll forward. Wang Ying, Baidu Group's vice president, mentioned in her August 14 share a judgment: merely completing tasks is no longer the core competitiveness; the industry's next phase is moving from task delivery to human-machine collaborative evolution. In product terms, this means an Agent should not return to square one after completing a task—a user-modified PPT is closer to real requirements than the initial draft, the Sedimented Skill (chéndiàn xià lái de Skill, meaning " Sedimented 技能" or "accumulated skills") can be directly invoked by the next task, and personal experience becomes organizational shared assets in the enterprise version. Once the cycle of "more usage, thicker context, more accurate results" starts spinning, it will be hard for latecomers to replicate.

However, it must be acknowledged that current Agents are still some distance from being "reliable." A detail easily overlooked in the National Industrial Information Security Center's evaluation report is that the average scores for content authenticity in Word and PPT among tested products were only 65 and 61.25, respectively, with some products still generating false data or hallucinations. Running through workflows is one thing; whether the output is trustworthy is another—this is a question all players must face next.

05

The Endgame Is Far from Set, but the Window Is Short

Despite the white-hot battle, this war is far from reaching its endgame.

Beyond the heat, the first obvious issue is retention. WorkBuddy's 20.97 million is a "visits" figure, not MAU; its DAU numbers vary widely across caliber (kǒujìng, meaning "metrics"): Tencent Chief AI Scientist Yao Shunyu mentioned "millions of DAUs" in a July interview, while Citigroup's research report estimates 20 million MAU and 13 million DAU. Launching new products with free subsidies and saturation advertising can easily boost curves; the challenge is keeping users engaged—Tencent itself has learned this lesson, with Yuanbao's early hype from advertising fading under more aggressive competitor follow-ups. Tencent Senior Executive Vice President Tang Daosheng explicitly stated in June that WorkBuddy is still in a strategic investment phase without commercialization KPIs, comparing it to "Tencent Meeting a few years ago": first scale, then find a model.

While commercialization models are still being explored, security—a must-answer question—has already arrived. After OpenClaw's explosion, security agencies monitored over 270,000 instances directly exposed to the public internet, with the National Internet Emergency Center and MIIT's vulnerability platform issuing risk warnings in March; multiple universities even banned installations. As Agents gain greater permissions—reading files, sending emails, executing code—the cost of any prompt injection or unauthorized operation far exceeds that of a chatbot giving a wrong answer. For enterprise clients, "can the company control the AI" is as important a purchasing reason as "can employees use the AI"—this is why all companies are piling on permissions, auditing, and privatized deployment capabilities in their enterprise versions.

There is also a broader question hanging in the air: When tech giants turn office suites into infrastructure for distributing Tokens, where will the space for startups be squeezed? Currently, a division of labor is taking shape—major companies dominate the general office sector, while startups delve into vertical business systems like ERP and CRM, developing value-added Agents deeply embedded in industry processes. The tripartite dominance of search, e-commerce, and social media during the PC Internet era may not repeat itself, as the pace of product iteration and data flywheel effects is much faster this time around, with a potential window of opportunity lasting only one or two years.

Looking back at this August, the near-simultaneous consolidation by four companies was no coincidence. Model convergence has pushed competition to the application layer, with office scenarios narrowing the competition at the desktop entry point. Ultimately, the battle for desktop entry points will hinge on one thing: whose Agents have more real work to do every day and can retain the context of each task.