Doubao is both honey and arsenic.

ByteDance is carrying out an increasingly clear initiative: to centralize all AI capabilities scattered across its various product lines onto the core platform of Doubao.

According to exclusive disclosure from *Intelligent Emergence*, ByteDance recently completed a team integration for its office AI products, with the TRAE and Coze teams fully merged into the Doubao system. Going forward, TRAE Work and Coze will be integrated with Doubao's work scenario capabilities; TRAE IDE and CLI will continue to be developed as programming product lines under the Doubao brand.

Next, ByteDance will also launch an independent AI office product "Doubao Work".

This has gone far beyond the scope of an ordinary team adjustment, especially since it is less than a month after the adjustment of Feishu. (Extended Reading: Feishu has been integrated into Doubao)

At ByteDance's all-staff mid-year meeting in August this year, CEO Liang Rubo put forward three key principles: top priority, thick main trunk, and long-term optimization, making it clear that Doubao will become the only core main trunk.

Now, this statement is being implemented at the organizational and product levels.

The AI capabilities previously explored separately by TRAE, Coze and Feishu are starting to converge towards Doubao.

Concentrating resources, teams and scenarios to focus on core targets is undoubtedly a huge advantage for ByteDance.

But on the other side of the coin, when all the company's AI chips are placed on one table, Doubao has instantly become the largest single-point business in ByteDance's AI strategy.

I. ByteDance starts to centralize its AI business layout

Looking back at the layout over the past few years, ByteDance's AI development has always been characterized by a strong "horse race" mechanism.

Feishu developed the Aily intelligent assistant, Coze built the Agent development platform, TRAE started from AI programming, and Doubao was responsible for verifying the large model and C-end assistant scenarios.

This horse race mechanism is highly consistent with ByteDance's usual logic: expand first, then screen; try and error first, then concentrate resources.

Relying on its huge engineering team, traffic pool and capital reserves, ByteDance opened four or five battlefields at the same time. It will continue to increase investment for projects that achieve good results, and quickly adjust direction for projects that fail to deliver expected outcomes.

But at the current stage of AI development, the Agent track has become an extremely capital-intensive heavy-asset consumption war.

Model training, computing power infrastructure, Agent toolchain, and ecosystem construction all require continuous massive resource investment. Internal mutual consumption will only lead to scattered combat effectiveness.

A more critical change lies in the competitive dimension.

I have always been curious about why new features such as remote control, virtual desktop, cloud computer, and side workbench have frequently appeared in the market in the past six months. When you look at them together, it is clear that they are no longer isolated functional points, but point to the same trend:

AI is evolving from simply answering questions to taking over tasks in real scenarios.

Under this trend, ByteDance has to accelerate its centralization process.

TRAE's programming experience, Coze's Agent ecosystem, Feishu's enterprise-level workflow, combined with Doubao's underlying model and C-end user scale, are being aligned to the same pace. In the past, each team searched for its own AI battlefield, and now Doubao has become the only destination.

This means that ByteDance's AI strategy has undergone a complete transformation: shifting from the horse race trial-and-error model to concentrating resources on building the core main trunk.

II. Doubao has become the aggregation of all AI capabilities

If you only look at the organizational structure, this adjustment can easily be misunderstood as the merger of several products.

But the real core change lies in the reconfiguration of underlying capabilities.

To enable an Agent to truly perform work tasks instead of humans, a single dialog box is far from enough.

It must be able to accurately understand user intent, call external APIs, break enterprise data barriers, and stably output results in complex business links. This means that an AI product must have both a "brain" and "hands and feet" at the same time.

The competition of chat products lies in how well they answer questions. The competition of work products lies in whether tasks can be completed successfully.

This is why ByteDance must integrate TRAE, Coze and Feishu into Doubao.

In the new product puzzle, the roles of each team have also changed:

Coze: Agent development and connection capabilities

TRAE: In-depth programming scenarios

Feishu: Organizational relationship, collaboration graph, workflow

Doubao: Model capabilities + the largest AI user base

ByteDance does not negate the previous explorations of all lines. TRAE IDE will continue to iterate, and Coze's development ecosystem will still be retained. They are no longer independent business lines, but start to reorganize around a unified work entry.

Therefore, the real goal of "Doubao Work" has become clearer.

It is far more than just an "AI version of WPS", nor is it simply embedding several office functions into a dialog box. It is more like ByteDance's first attempt to integrate models, Agents, tools and work scenarios into the same system.

This is a very large resource increment for Doubao.

But the real difficulties also start from here.

III. ByteDance has to step into its most unfamiliar battlefield

Over the past decade, ByteDance's strongest capabilities have always been rooted in the C-end market.

It finds high-frequency scenarios, obtains a large number of users, continuously improves matching efficiency through algorithms, and finally realizes commercial monetization. Douyin, Toutiao and even Douyin E-commerce are essentially extensions of this logic.

ByteDance is extremely good at leveraging C-end traffic, but it has rarely tackled hard problems in the deep water zone of the real B-end market.

C-end: high-frequency scenarios + traffic + algorithm matching, ByteDance's strongest field

B-end: security, permission, process, ecosystem, ByteDance's most unfamiliar field

But enterprise office follows completely different game rules.

The core assets of enterprises are hidden behind deep barriers: high-security level documents, strict permission control, attendance and approval processes, private knowledge bases, CRM, ERP. These data cannot be automatically crawled through traffic algorithms, let alone appear in a C-end assistant out of thin air.

Even Doubao, which has huge user advantages, has to adapt to the traditional enterprise software system that has been operating for decades.

At this intersection, Feishu is a natural interface for Doubao to penetrate into the enterprise world.

But having an interface is a completely different matter from being trusted and used for a long time.

Facing Microsoft Copilot, Google Workspace, as well as domestic players such as DingTalk and WPS that have been deeply cultivated for many years, what ByteDance lacks most is exactly the deep accumulation of enterprise relationships and the understanding of business workflows in specific industries.

On the surface, "Doubao Work" is to make AI office products, but in essence, it is to answer a very realistic question:

Can ByteDance complete the genetic transition from a traffic giant to an enterprise-level productivity service provider?

IV. Resource concentration is the biggest advantage

From the perspective of resource allocation, this is undoubtedly a highly strategic integration.

In the past, multiple teams repeatedly built Agent platforms and toolchains in different divisions. After the integration, R&D, technology, user traffic and operating costs are all gathered in one place instantly.

The most direct changes have appeared in Doubao's recent product updates:

Remote computer control

Windows virtual desktop

Cloud computer

Side workbench

More than 200 skills and connectors

This means that Doubao can recombine the capabilities that have been verified by different teams faster.

This capability reuse is especially important in the Agent era, because the competition of Agent products is increasingly like building blocks.

What really determines the iteration speed is not only how powerful the model itself is, but how many existing capabilities can be directly called.

In the past, to develop a new function, it might need to be built from scratch. In the future, a mature Agent system can directly combine existing capabilities.

This will bring the re-emergence of scale effect.

The more users there are, the more feedback you get. The more capability modules there are, the faster the product can be combined. The richer the connectors are, the more tasks can be completed.

This is a flywheel completely different from the traditional Internet traffic growth model.

The biggest significance of resource concentration for Doubao is not the increase in personnel. The key is that the previously scattered capabilities now have the opportunity to generate compound effects.

This is the possible dividend after ByteDance integrates all the capabilities accumulated from the horse race mechanism into one core main trunk for the first time.

V. The main trunk strategy pushes Doubao to the spotlight again

Problems also emerge at this point.

When Doubao is established as the main trunk of ByteDance's AI business, and all of ByteDance's chips are placed on this brand, the expectations and pressure it bears are also increasing simultaneously.

In the past, if a sub-product under the horse race mechanism failed, it could be regarded as a normal innovation cost.

But now, all product lines, teams and business expectations are tied to Doubao, and ByteDance's once most proud high-fault-tolerance innovation mechanism is failing.

After the trial-and-error space is compressed, Doubao is facing a battle that it cannot afford to lose.

This pressure will be reflected in every single indicator.

The underlying large model must remain in the first echelon, product experience iteration cannot have any lag, the Agent developer ecosystem must continue to thrive, B-end scenarios must win leading large enterprises, Feishu's collaborative benefits must bring real revenue, and the huge active user base on the C-end must not decline.

All these complex and interrelated goals will eventually be shouldered solely by the Doubao brand.

Doubao has finally evolved from a growth-oriented product to a strategic product.

The biggest difference between the two is that the former allows constant trial and error, while the latter has to constantly prove its value.

This is the downside of the advantages, the hidden risk.

The more concentrated the resources are, the greater the amplification effect will be when it succeeds. For ByteDance, this is actually a new organizational risk.

In the past, ByteDance's strongest advantage was that there was always the next promising project. Now, it has to prove that this single project can run the whole race successfully.

VI. What ByteDance needs to make up for is the long-term trust of the enterprise world

Most of ByteDance's proud R&D and product capabilities in the past were cultivated in C-end scenarios, pursuing extreme recommendation efficiency, traffic monetization and agile iteration.

After entering the enterprise market, what ByteDance needs to accomplish is far more complex than developing an office Agent.

It needs to transform the product capabilities it is best at in the past into system capabilities that enterprises can rely on for a long time.

These two capabilities seem similar, but there is a huge gap between them in essence.

C-end products pursue speed, ease of use, high frequency and continuous iteration. Enterprise software pays more attention to stability, security, permission, ecosystem and long-term service.

Especially when AI begins to enter the core work of enterprises, what enterprises purchase is no longer just a software account.

What they purchase is certainty.

If a financial Agent occasionally gives a wrong answer, it is acceptable. But if it modifies financial data incorrectly, the problem will be completely different.

This is why ByteDance must re-understand the enterprise market.

The more powerful the Agent is, the more important the permission control is. The deeper AI penetrates into the business, the more important data governance is. The more complex the task is, the more important system stability is.

From this perspective, the significance of Feishu, TRAE and Coze to Doubao is not simply adding several functions.

They respectively represent several layers of infrastructure that enterprise AI needs to face: organization, collaboration, tools, development, permission, process.

If these capabilities can be truly connected, ByteDance can transform its strong product iteration capabilities into system capabilities that enterprises rely on for a long time.

This is more difficult than developing a hit AI product.

VII. ByteDance is stepping into the Doubao era

From a long-term perspective, this adjustment releases a very clear signal: ByteDance's strategic pillar is undergoing a historic shift.

In the past decade, ByteDance's business empire had an irreplaceable main axis: content.

Toutiao distributes information, Douyin distributes short videos, algorithms complete the efficient matching of traffic, and finally realize business closed loop through advertising and e-commerce. All of ByteDance's derivative businesses in the past are essentially transfusing blood or extending to this main axis.

Now, AI is being pushed to the position of the next main axis, and Doubao is the selected carrier.

Past main axis · Content: Toutiao + Douyin + algorithm, everything revolves around content and traffic

Future main axis · AI: Doubao + TRAE + Coze + Feishu, everything is reorganized around Doubao

The integration of TRAE and Coze is just the first move, the interconnection between Feishu and Doubao is the second move, and the upcoming launch of "Doubao Work" will push this organizational change directly to the front end of products.

The most important thing ByteDance has done this time is not incubating a new AI office product. It has finally made up its mind to build a solid main trunk for its AI strategy.

This main trunk was once named Douyin. In the future AI world, its name is Doubao.

Beyond the Layout:

The Internet used to solve the problem of connection. AI is starting to solve the problem of organization.

The former makes information, content and goods flow. The latter may enable capabilities, tools and workflows to be truly reorganized.

ByteDance has proved how good it is at doing the former thing. What Doubao needs to answer next is the latter.

This article is from the WeChat official account "Beyond the Layout", author: Huahua, published with authorization from 36Kr.