After OpenClaw gained popularity in March this year, major domestic tech companies rapidly launched similar products, leading to severe duplication of efforts and functional overlap in the office agent market. In July, Alibaba integrated QoderWork, Wukong, and MuleRun into Qwen Office, under the leadership of DingTalk CEO Chen Yusen; Tencent restructured its QClaw-related businesses under Cloud Product Division Six. Around the same time, mainstream apps such as DouBao, Qwen, and Tencent Yuanbao simultaneously shut down their agent marketplaces. Major companies have shifted from expanding broadly to consolidating their efforts, with industry consensus forming around integrating traffic entry points and AI agents—but how to get users to pay remains a challenge.

Author and source: Shixiang

“Previously, the biggest risk for agents was betting on the wrong direction; now it’s redundant development,” said someone close to the AI industry to “Shixiang”: “With so much capital and computing power available, horse racing isn’t a sustainable strategy.”

In the past, major domestic companies rushed to capitalize on the OpenClaw trend, flooding the market with a batch of repackaged agents.

Today, as "Agent shells" litter the landscape, major companies are preparing to move in.

On July 2, Alibaba announced the integration of QoderWork, Wukong, and MuleRun into a new product called Qianwen Office, overseen by DingTalk CEO Chen Yusen.

On July 20, Tencent transferred the relevant business and team of the QClaw Product Center to Cloud Product Division Six, which oversees WorkBuddy.

These two changes also marked the end of the OpenClaw trend.

The large tech companies' office intelligent agent matrix was a panic response following the rapid rise of OpenClaw. In March this year, after OpenClaw swiftly topped GitHub’s all-time star leaderboard, Tencent’s Qclaw, ByteDance’s ArkClaw, and Alibaba’s JVS Claw quickly followed suit. Within a short time, a flood of wrapper agents based on OpenClaw’s encapsulation or compatibility were rapidly produced.

The cost of rushing to market was that various products within the company had highly overlapping features, all focused on white-collar office scenarios, differing only in their entry points and traffic sources. For example, MuleRun entered through Alibaba Cloud, QoderWork through the desktop client, and Wukong through DingTalk collaboration.

At Tencent, WorkBuddy is developed by Tencent Cloud, while Qclaw is developed by Tencent Computer Manager—users themselves struggle to distinguish between them.

At ByteDance, at least four agents were developed internally—Feishu Miaoda OpenClaw, Volcano Engine ArkClaw, Kozu OpenClaw, and Trae Work—with multiple teams independently customizing OpenClaw for productization.

Multiple product lines in competition offer the advantage of abundant ideas and creativity, but the downside is that resources are too dispersed, making them suitable only for defense, not for large-scale battles. OpenClaw once iterated through three versions in three days, breaking through the defenses of major companies as a weekend project by a retired programmer.

The integration of Tencent WorkBuddy with Qclaw, and Alibaba’s integration of Qwen Office, both fundamentally carry a sense of damage control.

In early July, a more symbolic industry event occurred simultaneously: major large model applications such as ByteDance's Doubao, Alibaba's Qwen, and Tencent's Yuanbao shut down their agent marketplaces and custom agent creation portals almost at the same time.

Major companies have rapidly shifted from broadly expanding their AI applications to focusing their efforts—a trend that has become an industry consensus.

The era of Wukong on DingTalk has come to an end.

The surge of Tencent Cloud’s WorkBuddy indirectly highlights the issue faced by Volcano Engine’s head, Xiao Mo, who, upon seeing OpenClaw earlier this year, redirected all efforts and resources toward it. As a result, Trae Work was launched months later than competitors, allowing WorkBuddy to dominate the market; with the entry point lost, opportunities have steadily diminished.

This statement comes from a ByteDance employee. The conclusion is clear: major tech companies are integrating disparate agent products into workspaces, a trend that is becoming dominant in securing the core entry points of the future.

The integration of Qwen Office is similar. On the surface, it appears to be the merger of three Agent products, but behind it lies the continuation of the ATH business group’s integration since March. Meanwhile, DingTalk is also using Qwen Office to redefine itself.

From a timeline perspective, on March 16, Alibaba’s ATH division consolidated five entities—Tongyi Lab, MaaS business line, Qwen Business Unit, Wukong Business Unit, and AI Innovation Business Unit—into one. Just 24 hours later, Alibaba launched Wukong, an AI-native work platform, solidifying its strategy of combining a traffic gateway with an agent-based work platform.

The personnel changes in June accelerated the integration process, with Chen Yusen, founder of MuleRun, succeeding Wu Zhao as CEO of DingTalk. Previously, he led an internal startup at Alibaba Cloud, delivering the only product within Alibaba to achieve actual commercial data in the Agent space.

Chen Yusen is leading DingTalk, which is also seen as Alibaba putting its best AI product expert in charge of the overall strategy.

Just one week into his tenure, Chen Yusen announced a comprehensive restructuring of five departments, with Wukong and MuleRun merging to form a new Wukong team. Shu Junliang, former CTO of MuleRun, has been appointed as the head, effectively placing MuleRun’s technical team in charge of the entire Agent product line, marking the end of the original Wukong team under Wu Zhao’s leadership.

Meanwhile, six channels—direct sales, telesales, service providers, pre-sales, delivery, and customer success—have been consolidated into a single Customer Development Department, ending the previous situation where Kung Fu, DingTalk, and MuleRun had separate sales and delivery teams.

By the triple integration announcement on July 2, Qwen Office, built on the QoderWork foundation and integrating Wukong and MuleRun, is organizationally closer to the DingTalk product line.

The positioning of DingTalk's AI application carrier has been further strengthened.

According to the "Q2 2026 China Office Agent Platform Market Insights Report" released by Analysys, as of June this year, Tencent WorkBuddy had 20.97 million monthly visits, ranking first in the industry, while Alibaba's QoderWork had 7.88 million.

Alibaba lacks social virality and C-end customer acquisition capabilities; ByteDance has Douyin and Feishu, while Tencent has WeChat and QQ. Alibaba’s core advantage lies in the B-end market with DingTalk.

DingTalk’s base of over 26 million enterprise organizations defines Qwen Office’s product roadmap. Unlike WorkBuddy, which targets native internet users, Qwen Office will focus more on delivering tangible outputs and efficiency gains to rapidly capture the B2B market. Combined with Alibaba Cloud’s bundled sales, the package of Qwen Office, cloud servers, and model services enables Alibaba to deliver a competitive combination that rivals find difficult to replicate.

For DingTalk, the department has repeatedly attempted to integrate Agent capabilities into the product. After this integration, DingTalk will evolve from an AI tool to an operating system, enabling enterprise employees to directly invoke Qwen Office and embed AI into their workflows.

Integrating with DingTalk is also about saving time. For users, the cost of replacing AI tools like WorkBuddy, Alibaba Tongyi, and Coze when deeply embedded into enterprise CRM systems, DingTalk/Feishu workflows, and code repositories is extremely high—user data, habits, and relationship networks are all deeply rooted here.

This is also why Qwen Office must be integrated. By enabling DingTalk to directly launch Qwen Office for writing weekly reports and pulling data charts, we aim to build user muscle memory and secure the entry point for desktop productivity.

WeChat Work didn't become Tencent's AI agent partner.

In this round, major tech companies have collectively shifted their focus in the office agent space, with AI integrated with traffic entry points and organizational structures becoming a standardized approach.

Alibaba, ByteDance, and Baidu have all adopted a bundled strategy combining traffic entry points with AI agents. Baidu’s GenFlow Agent is directly embedded into its cloud storage and document library, while Qwen Office is deeply integrated with DingTalk.

ByteDance has taken a more aggressive approach, with all four agent products centered around Feishu.

Feishu Miaoda OpenClaw is natively integrated within Feishu, ArkClaw is seamlessly integrated via Feishu Channels or messaging channels, Kozu OpenClaw connects to the Feishu ecosystem through plugins, and Trae Work enables integration with Feishu Docs and Multi-dimensional Tables via external app authorization links. Feishu serves as the common destination and traffic hub for all of ByteDance’s agent products.

Tencent, on the other hand, took a different path.

WorkBuddy is not tightly integrated with WeChat Work; instead, it actively supports integration with Feishu and DingTalk. Users can remotely control WorkBuddy through IM platforms such as WeChat, WeChat Work, QQ, Feishu, and DingTalk.

In Tencent’s view, agents should aim for low barriers to entry rather than control, serving as independent work entry points. The organic growth of WeChat and social platforms is the true traffic engine; the customer retention effect of Enterprise WeChat is optional.

By integrating over ten leading models such as HunYuan, DeepSeek, GLM, Kimi, and MiniMax, and connecting cross-platform with office software, WorkBuddy enables users to remotely control their computers for work.

Regarding the positioning of WeCom, it is more akin to an ecosystem channel rather than a core entry point.

Whether an organization is integrated is another important dividing line.

Alibaba has assigned Chen Yusen to lead its office Agent product portfolio, while Tencent has assigned QClaw to WorkBuddy—everyone is choosing to consolidate, combining fragmented capabilities into a unified force.

The Baidu case clearly illustrates the point. Since 2026, Baidu has shifted its strategy from dispersion to integration, establishing the Personal Super Intelligence Group (PSIG) and incorporating traffic entry points such as Baidu Wenku and Baidu Netdisk into it. PSIG has since launched the cross-platform universal agent platform, GenFlow Agent.

The goal is clear: enhance business weight through an independent business unit, separate C-side assets from MEG as a growth engine, and integrate all validated assets to ensure the creation of an AI super gateway.

An agent is not a single technological product; it requires end-to-end capabilities including model performance, engineering, product design, distribution channels, and customer sales. When these capabilities are dispersed across different business units, it becomes difficult for management to enforce a unified vision and allocate resources effectively.

User retention data for OpenClaw products further underscores the necessity of integration. Growthbox data shows that nearly 70% of users decide to abandon their Agent within two weeks of use.

But bytes are an exception.

Its four agents belong to the Volc Engine BU, Flow, and Feishu BU, share the Seed foundation, but operate independently within each BU.

Behind the scenes, Feishu and VolcEngine are both first-tier business units with tens of thousands of employees, each operating with independent accounting, independent decision-making, and direct reporting to the CEO. This is entirely different from Alibaba and Tencent, which integrate agent teams of around a hundred people.

To strike a balanced approach, we’ve adopted a unified foundation with divergent applications. DouBao provides a standardized training framework, while each product captures a distinct positioning: Feishu Miaoda is a product add-on, ArkClaw is an independent SaaS, Kozu is an ecosystem platform, and Trae Work is an AI productivity assistant designed to claim users’ desktops—each with its own strategic mission.

AICoding becomes a paid high ground

The current challenge for large tech company office agents is whether users will pay for them.

Take Qwen Office as an example. Backed by DingTalk, which has a massive user base, its paid conversion rate remains low. At DingTalk’s 10th anniversary event in 2025, Wu Zhao revealed that DingTalk serves over 26 million enterprise organizations, with 79% of the 5,191 A-share listed companies using the platform—yet the number of paid organizations does not exceed 200,000, resulting in a conversion rate of less than 1%.

In other words, DingTalk's core user base consists precisely of small and medium-sized enterprises with low AI adoption rates and low willingness to pay, which contradicts the product logic of AI agents that requires deep integration into workflows, long-term usage, and data accumulation.

In the AI-focused coding track, domestic tech giants also lack an advantage.

In the May 2026 Code Arena third-party programming leaderboard, Alibaba's Qwen3.7-Max scored 1541, surpassing models such as GPT-5.5, Gemini-3.5-Flash, GLM-5.1, and Kimi-K2.6, and ranking second globally among large model providers, behind only the Claude series.

腾讯WorkBuddy has initially achieved success with its individual paid tier. The Professional Plan for individuals is priced at RMB 58 per month, offering a monthly allocation of 2,000 points with no usage frequency restrictions. According to Tencent's first-quarter financial report, the retention rate for paid users of WorkBuddy exceeds 80%.

On the other hand, core coding users prefer to pay $100 to $200 per month in API fees to subscribe to Claude Code, accessing it through proxies and cross-border payments, rather than using domestic agents for free.

The model's capabilities are too far apart; Qwen writes code that's too verbose—for the same requirement, Claude uses 400 lines, while Qwen gives you 1,000 lines." A user told "Shixiang" that although overseas products are expensive, their capabilities justify the price.

This is not an issue of model benchmark scores, but rather a significant gap between the model’s capabilities and the Agent product’s capabilities. Core users require the Agent’s deep task capabilities, such as cross-file architecture refactoring, contextual reasoning on legacy code, running tests, analyzing error logs and stack traces, and modifying code accordingly.

An insider from the AI industry commented on "market sentiment," noting that the market is focusing on Anthropic not for ordinary conversations, but for higher reliability, stronger task completion rates, and greater willingness to pay within enterprise workflows.

Currently, domestic models lack high-quality, top-tier open-source code for coding tasks; their products primarily aim to reduce costs and improve efficiency for programmers rather than fully replace human developers, which leads experienced engineers to avoid spending time on paid subscriptions.

The response from large companies has been to impose internal bans.

In early July, Alibaba required all employees to uninstall all Anthropic products by July 10, including the Claude Code, Sonnet, Opus, Fable, and other model series, replacing them uniformly with the company’s self-developed intelligent agent programming platform, Qoder.

Although this notice is framed with security considerations in mind, the underlying logic is that businesses need administrative measures to create demand for their own products.

For Qwen Office, the product must show enterprises not how intelligent AI is, but how much money AI can save them. These metrics must be consolidated into PPTs so decision-makers can see reduced labor costs and shortened delivery cycles.

In the next AI cycle, it won’t be about how many agents a company has built, but which product makes users unable to let go.