In just two months, on the personal Agent table, players have Dense Bright Cards (intensively revealed their cards).

Meta's Muse launched and then receded, with daily net new users plummeting from a September peak of 198,000 to 74,000, a 62.4% drop. Alibaba's 'Qianwen Go!' registered user intent by successfully processing a flight delay claim for just RMB 70. Tencent's WeChat Xiaowei remains in grayscale expansion.

The three players have different styles and progress, but at least they are all at the table.

Meanwhile, ByteDance's earliest internally tested personal assistant, Xiaodou (Spell), remains 'in internal testing' with no announced launch date or observable cases.

ByteDance, which has long held a strong hand, has yet to play its card, which contradicts external perceptions of the company.

ByteDance was once China's most product-savvy company: In the Graphic and textual information flow (text-and-image feed) era, Toutiao redefined distribution with its recommendation engine, and TikTok replicated the same product mechanism globally. In short videos, Kuaishou led, but Douyin overtook it with swipe experiences and music templates. Even tools like CapCut became the default for short video editing.

This is an App factory accustomed to late-mover advantage, disrupting markets with products and algorithms. However, in the Agent battle, ByteDance's efforts on both ends have been underwhelming. Its personal assistant was preempted by 'Qianwen Go!,' while the workplace market was first washed by WorkBuddy.

The reason may not lie in the products themselves. ByteDance's past core strength lay in organizing products, algorithms, and traffic into an efficient machine for rapid cold starts, retention, and scale. Agent competition, however, hinges on model ceilings, ecosystem empowerment, and self-revolution of business models—all of which demand more from ByteDance's capabilities.

I. Personal Agents Await the Wind

Inaction implies greater consideration.

Personal Agents are ByteDance's strongest suit: high-frequency, unselective user base, and a network effect that scales to dilute computational costs. Morgan Stanley estimates Muse's cost per user per month ranges from $3 to $130. Muse thrived early in Meta's digital content ecosystem.

Douyin and Doubao's user pools offer similar potential. For ByteDance, this is a battleground to replicate its mobile internet expansion strategy. Fully activating hundreds of millions of C-end users in the AI era holds immense opportunity.

Looking back, ByteDance has been 'blocked' twice in the personal Agent arena.

The first time was late last year. The initial Nubia Doubao phone attempted to bypass app blockades by implementing 'decision-making for users': in shopping, it could auto-compare prices, search, and place orders across apps, with users only needing to complete payment. The phone launched to peak popularity, even commanding a RMB 5,000 premium in the secondary market.

This GUI Agent-based 'accessibility interface' approach relied on system-level permissions to read screens and simulate clicks—essentially using users' eyes and fingers to force open other apps' gates.

No incumbent would tolerate this. Within a week, starting December 2, users operating WeChat via AI triggered 'environment anomalies' and forced logouts. On December 5, Doubao suspended financial operations. By December 10, some Alibaba apps lifted login restrictions only if Doubao disabled related operational capabilities.

Having 'hit a wall,' ByteDance adopted a 'courtesy first, force later' approach in its second clash. On September 14, with the second-gen Doubao phone (Nubia NaviX Ultra), it launched the Screen Automation Operation Statement Protocol (SAEP), giving all apps a 30-day window to opt-in or out of Doubao's proxy operations, with fine-grained permissions and high-risk actions restricted by default.

By October 14, the grace period ended, and 153 major apps nearly all rejected the protocol—WeChat, Alipay, Taobao, Meituan, Pinduoduo, Didi, and Xiaohongshu were absent, as were DeepSeek and Kimi. The next day, ByteDance reversed to a 'default allow, unless explicitly denied' rule.

Thus, both paths for the nascent personal Agent product 'Xiaodou' are challenging.

Externally, core interests block cross-app universal assistants. Allowing ByteDance's Agent into their ecosystems would mean surrendering core user behavior data and transaction entry points to a competitor—a non-starter for most platforms.

Internally, Xiaodou's potential space may shrink to the Douyin ecosystem, acting as a 'ByteDance-style life assistant.' However, if the Agent bypasses recommendation feeds to complete transactions directly, it undermines ByteDance's core advertising revenue model.

To make this conflict tolerable, on August 10, wine and tourism orders processed via Doubao faced a higher independent fee rate (~12%) than Douyin's main platform (8%), serving as a warning shot to internal and external ecosystems.

This commercial dilemma seems to make the typically aggressive ByteDance hesitant.

Qianwen Go! dared to play first, not because Alibaba lacks advertising baggage. Taobao and Tmall also rely on transaction commissions and merchant ads, but Agent-driven consumption occurs within the ecosystem—the more successful the Agent, the higher the commission efficiency and the less erosion of ad revenue. ByteDance's closed loop (closed loop) is an 'attention loop': every minute the Agent saves for users means one less minute of ad inventory for Douyin.

With Xiaodou, the main Doubao app, and Doubao Work all consolidated under Zhao Qi, and Xiaodou—reportedly in development since April—still unseen, the delay likely stems from ByteDance's hesitation over its business model.

II. Sharp Tools, Yet Unfinished Work

ByteDance's workplace Agent front appears gentler for now.

When 'Doubao Work' launched on August 25, WorkBuddy, which debuted in March, had led for nearly six months, with PC monthly visits reaching 20.97 million in June—industry-leading. As WorkBuddy integrated with Tencent Meeting, Tencent Docs, WeChat, and Enterprise WeChat, Agents gained access to richer contexts, improving practical effects and user mindshare.

ByteDance, known for late-mover disruption, this time attempted a 25-day lightning integration to catch up.

On July 30, the Feishu product team merged into Doubao, with Xie Xin reporting to Zhao Qi and sales/marketing integrating into Volcano Engine. On August 24, TRAE and Kouzi teams joined the 'Doubao family.' On August 25, Doubao Work officially launched, integrating with the Feishu ecosystem.

By September 15, Feishu 8.0 debuted, featuring a side navigation bar for direct access to Doubao Work, supporting reading of group chats, documents, and other Feishu functions. Basic office and collaboration features were consolidated into Doubao Work, deepening their bidirectional integration and achieving full product-level connectivity.

This highly efficient integration responded to inevitable era shifts. Product-wise, no AI-native app has yet fully replaced last-generation tools. Organizationally, it resolved internal friction—previously, Feishu, Doubao, TRAE, and Kouzi operated independently, with fragmented entry points and data silos, forcing enterprise clients to engage with multiple ByteDance solutions.

But resolving internal friction is about 'sharpening tools,' not necessarily 'accomplishing tasks.' At the ecosystem level, ByteDance faces deeper structural choices.

Workplace Agents represent the enterprise-side implementation of Token economics. By productizing underlying Tokens, they enable pricing and drive consumption, generating service revenue wherever Tokens are spent.

Under 'equal user demand generating equal Token consumption,' open multi-model and proprietary model approaches may seem equivalent in revenue scale but diverge sharply in technical evolution and commercial competitiveness.

Open ecosystems excel in rapid engineering iteration and traffic siphoning. Take Tencent's WorkBuddy: in one month, it became one of the first to integrate Space Bunny (anonymous) and Step 5 Preview from StepFun—offering high freedom to switch and configure models without complex API access, empowering users in the AI era.

Under this route, open platforms abandon single-foundation-model binding, focusing instead on engineering orchestration, connector ecosystems, and workflow encapsulation. The more open the platform, the more freely the interfaces, the more Tokens flow—and with them, the most active users.

Doubao Work, choosing closure, bets on data loops and foundation-model iteration.

By adhering to a 'closed suite' route with only its in-house Seed model and Feishu contexts, Doubao Work sacrifices short-term appeal to users with diverse needs. The trade-off: locking all Token consumption and interaction scenes into its Seed model, ensuring full real-data reflow (reflow) to continuously feed and refine its foundation model via evaluation feedback.

From this perspective (perspective), Doubao Work serves Zhang Yiming's strategic will to 'exhaustively improve model capabilities,' making the workplace Agent a training ground for the Seed model.

Meanwhile, enterprises purchasing workplace Agents essentially buy Tokenized services for organizational data—migrating their structures, permission systems, documents, approvals, and collaborations wholesale.

Market once saw 'Doubao's 345 million MAU' as ByteDance's key asset. Yet Doubao's massive MAU comprises C-end personal users who chat with AI—far removed from organizational spending.

Consider another metric: DingTalk has 171 million MAU, directly undertake (inheriting) Qianwen's office capabilities, while Feishu has ~30.19 million MAU—about one-sixth of DingTalk's. WeChat Work, backed by WeChat's social graph, can penetration (penetrate) without heavy reliance on its own collaboration platform, leveraging the WeChat-Enterprise WeChat relationship chain instead.

Granted, ByteDance isn't without opportunities. AI Product Rankings' July data shows TRAE Work at 2.66 million monthly visits, Kouzi at 2.17 million (+62% MoM). But what delights developers may not suit ordinary workers. A source close to TRAE Work said, 'It just copied the old IDE approach, leaving the product nearly stagnant in H1.'

AI's convergence is forcing tool evolution, where unified, seamless interactions promise predictable opportunities.

In the AI era, 'product kingship' may no longer mean cramming complex buttons into interfaces. Instead, it means using ultimate (ultimate) foundation models and engineering to keep complexity internal and deliver 'one-click' simplicity to users.

III. Technology-First on Both Fronts

Viewing ByteDance's personal and workplace fronts together, it's clear the company is shifting from its familiar product-first approach to a longer, costlier technology-first one—key to its slowed pace.

Past internet products relied on smooth interfaces, stronger recommendation algorithms, and higher growth efficiency. But for Agents to truly complete tasks for users, model reliability becomes core. Users may let Agents fetch info but hesitate to let them pay, approve, or handle contracts—ceding partial decision rights to AI.

Thus, Zhang Yiming set a goal to 'exhaustively improve model capabilities.'

Organizational adjustments signaled this shift: in February 2025, Wu Yonghui joined from Google DeepMind; in October, Zhu Wenjia reported to Wu, establishing Seed as the top research priority. Conversely, Qiao Mu exited, Yang Jianchao took a leave, Zhou Chang from Tongyi Qianwen succeeded him, and Li Hang became an advisor.

Research-focused leaders ascended, product veterans stepped back, and data/reinforcement learning departments were reinforced. New competitive metrics emerged.

Yet foundation models demand long-term investment—data, training, reinforcement learning, and post-training require time, compute, and talent. Unlike internet products, which validate returns through rapid launches and trials, technology-first approaches raise ceilings but are slow and costly.

After tilting resources toward technology, product consolidation lagged. Doubao Work's entry points remain unclear—desktop, Feishu sidebar, and Doubao App all offer access.

Beyond existing office users, ByteDance seems to want to convert Doubao's personal users into office users, but multiple entry points fragment R&D, operations, and enterprise services while blurring user perception. Is this a personal assistant, an office suite, or a Feishu AI enhancement?

Entry points for workplace Agents aren't just buttons—they determine where organizational structures, permission systems, documents, and collaborations ultimately reside.

In contrast, Qianwen explicitly separated Qianwen Office from its original entry point, targeting B-end users. According to Alibaba's September 2026 disclosure, when Qianwen Office surpassed 30 million users in its debut month, enterprise users accounted for over half (>50%).

From C-end Xiaodou to B-end Doubao, ByteDance's Agent fluctuations reflect an inevitable clash between traditional business models and new-era technological laws.

In the traffic era, ByteDance was an unstoppable 'efficiency machine,' distributing attention via algorithms and monetizing through ads. In the Agent era, the core competition shifts to model certainty, ecosystem openness, and deep organizational data trust.

The Agent war is a composite test of product, organization, and technology—a classic 'barrel effect.' ByteDance must prevent Agents from eroding Douyin's ad inventory while pushing closed-loop suites to persuade enterprises to migrate, even making products bear model evolution tasks.

Products and models must Mutual flywheel (act as mutual flywheels) and rotate sequentially. Beyond a clear entry point, an organization that unifies technology, products, and scenarios is essential. Technology raises ceilings, but products still decide when the flywheel starts.