AI Agent Race Ends: Giants Unify Entry Points, Battle for Rule-Setting Power

From Multi-Track Racing to Unified Entry Points

In the first half of 2026, each of China's four major tech companies ran parallel internal "horse races" across three fronts: desktop clients (WorkBuddy, QoderWork, DuMate, TRAE Work), IM-integrated assistants (Wukong in DingTalk, Feishu Aily, Enterprise WeChat "Big Circle", Ruiliu), and browser-based agents (Kimi WebBridge, Doubao virtual desktop, companion-style agents). This multi-track approach reflected genuine uncertainty about where AI agents should enter enterprise workflows.

August 2026: Synchronized Convergence

Within a single week in July–August 2026, all four companies abandoned the multi-track strategy:

Alibaba (most decisive): On August 3, merged QoderWork, Wukong, and MuleRun into QwenWork (千问办公) under DingTalk CEO Chen Yusen. The unified system combines local execution (QoderWork), 7×24 cloud long-running tasks (MuleRun), and organizational context via DingTalk. Positioning: enterprise-grade AI productivity platform, not a consumer tool; short-term MAU not a KPI.

Tencent (lightweight consolidation): Starting July 20, folded QClaw into WorkBuddy; launched creative agent Miora on July 22; introduced "human-AI co-writing" with Tencent Docs on July 30. No new master product — instead, WorkBuddy becomes the scheduling hub for Yuanbao, ima, QQ Browser, Enterprise WeChat Big Circle, and Miora, all built on the same agent framework.

ByteDance (organizational restructuring in three steps): July 30: Feishu product team moved under Doubao, reporting to Doubao head Zhao Qi; sales folded into Volcano Engine. August 24: TRAE and Coze teams merged into Doubao. August 25: Doubao Work released. Feishu's six years of documents, messages, org relationships, and permissions become the "key" for agents entering enterprises — Feishu downgraded from independent collaboration product to infrastructure pipe.

Baidu (slow start, aggressive expansion): DuMate fully opened March 22; internal agent dodo merged August 6; August 27 connected multiple Baidu products and released 15 industry suites with 96 professional skills. 150 iterations in five months; MAU month-over-month growth of 1,063.79%, becoming the fastest-growing product on AI office leaderboards alongside WorkBuddy.

Why Converge? Three Forcing Factors

Fragmented experience: Users needed three different apps from the same vendor to complete one continuous task (e.g., Wukong in DingTalk, QoderWork on desktop, MuleRun in cloud), forcing users to decide "which tool for this job" — directly hurting retention and reputation.

Resource dispersion: Three teams duplicating work diluted compute, data, and engineering talent. In 2026's intense foundation-model arms race, splitting firepower across three fronts ceded the main battlefield. ByteDance moving Feishu sales to Volcano Engine and Alibaba unifying talent under QwenWork both reallocated resources from "reinventing wheels" to the primary track.

Brand fragmentation: Searching "Alibaba AI office" returned three product names, preventing brand asset accumulation. QwenWork's launch site stripped vague terms like "platform, operating system, super entry" and listed only concrete capabilities: IM writing, Office generation, web generation — a "de-bubbling" of brand perception.

Analyst firm Analysys (易观) Q2 2026 report provided external validation: total desktop AI office agent visits in China exceeded 60 million in June, with WorkBuddy at 20.97 million — more than the second and third combined. Market had already chosen; marginal returns from continued experimentation approached zero.

The Real Contest: Three Deep Assets

Beyond the visible "entry battle," the convergence reveals competition for three structural assets:

Organizational context: The scarcest asset. WorkBuddy binds WeChat/Enterprise WeChat; QwenWork binds DingTalk; Doubao Work binds Feishu; Dazi connects Ruiliu. All aim to weave the enterprise's "how work gets done" memory — chats, documents, meetings, permissions, business-system data — into their own network. Atlassian's Teamwork Graph already proved: feeding context via graph structure yields +44% accuracy and −48% token consumption — context is not icing, it's a generational gap.

Business processes: Agent value isn't "writing text" but capturing intent, orchestrating tools, and writing results back into approval flows and CRM. QwenWork integrates 25 DingTalk IM capabilities (attendance, approval, meetings, docs); WorkBuddy's Agent Suite natively connects Tencent Docs, WeDrive, LeXiang. They're fighting for "workflow ownership." Digital economy expert Luo Rentong (骆仁童) pins it: the entry battle has never been about model parameter count, but about "workflow ownership."

Token distribution rights: The most overlooked, most lethal layer. When documents, meetings, approvals become token-consumable streams, whoever defines which scenario consumes whose tokens, calls which model, and routes through which billing chain holds the "utility metering rights" of the AI era. Tencent's SkillHub launched SkillPay, unifying skill distribution, agent invocation, and payment in one chain; Volcano Engine's 49.5% share tops China's MaaS market — both are staking claims on token-distribution infrastructure.

Costs of Convergence

Fewer user choices: During the racing phase, 17 desktop products competed; users could vote with feet. Post-convergence, each ecosystem offers one "official answer," artificially raising switching costs — use DingTalk, hard to switch to Feishu; adopt WorkBuddy's Tencent Docs loop, accept its scheduling logic by default. Market shifts from "hundred flowers" to "four kingdoms."

Innovation may slow: Internal racing, though wasteful, provided redundancy for innovation. Teams facing "lose and disband" bet on radical approaches. Single-track convergence compresses trial space; product forms homogenize — all become "one chat box + skill pile + scheduling hub." Organizational cost: DingTalk and Feishu, once demoted to pipes, simultaneously lose direct pricing power over customers, the liability interface for AI output errors, and the option for independent financing/spin-off. The reversal cost of organizational downgrade is extremely high.

Long-term, convergence can benefit users: fragmentation's flip side is paying, learning, and migrating across four non-interoperable systems. Unified entry lowers barriers, turning agents from geek toys into "out-of-the-box" productivity. The key is whether giants preserve open interfaces so users can take their data with them.

Verdict: Entry Is the Stage; Rule-Setting Is the Endgame

When WorkBuddy, QwenWork, Doubao Work, and Dazi all become "unified entries," their differentiation will no longer be features but rules — what data enters, which tasks auto-run, which model gets called, what billing standard applies, who bears liability for errors. The entry is the visible battlefield; rules are the invisible wall. Whoever defines the default options collects the tax on the next generation of SaaS.

Tencent's play: ecosystem binding — stitch WeChat, Enterprise WeChat, Docs, Meetings into an "AI operation layer," turning others' tools into OS components.

Alibaba's play: pragmatic consolidation — end the race, fix the main product, deliver a monetization plan, then talk ecosystem.

ByteDance's play: asset reclassification — move Feishu from profit center under the AI main line and cloud main line.

Baidu's play: capability aggregation — leverage Search, Miaoda, Famou breadth to make DuMate an independent entry unattached to any IM.

Four paths, one exam question. 2026 is widely labeled "China's Enterprise Agent Entry Definition Year." A pattern emerges: scenario assets closest to foundation-model capability radius — documents, meetings, code, customer service — will be "subjectively expropriated" first; moving six months late equals one voluntary devaluation.

This holds for every enterprise and individual using office software. The interface you open daily is becoming a tax collection point; your only move, before being piped, is to decide whether you are the grid or the appliance.

The multi-track horse race has ended, but the real competition is just beginning.