Manus has landed on the shore, but Agent hasn't won yet.

General or vertical? Should one focus on model capabilities or engineering capabilities? Should the market target be domestic or overseas? Manus has made the most pragmatic choice.

At the beginning of December, when asked about the most pessimistic scenario during an interview, Ji Yichao, the co-founder of Manus, replied, "It might go bankrupt next month." However, more than 20 days later, Manus was acquired by Meta for billions of dollars, marking an exclamation point in the year 2025, a year of rapid advancement in AI.

Media reports indicate that the deal was reached at lightning speed within ten days. Manus will continue to operate independently, and its founder, Xiao Hong, will serve as a vice president at Meta. This acquisition came just nine months after Manus launched its product.

Within nine months, Manus experienced a remarkable journey. In March 2025, Manus became an instant hit upon its launch, showcasing the ideal form of an Agent: with just one command, it can automatically call various tools to complete tasks such as writing code and planning trips. Its invitation codes were even resold for up to 100,000 yuan.

Meanwhile, Manus has also been mired in controversy, with many questioning it as the "king of shell wrapping" driven by marketing. In July, Manus laid off a large number of employees and relocated its headquarters to Singapore. At that time, there were rumors that Zhang Yiming of ByteDance had offered $30 million to acquire Manus, but Xiao Hong declined.

In December, Manus revealed that its Annual Recurring Revenue (ARR) had exceeded $100 million, making it the fastest-growing startup globally to reach $100 million in ARR from scratch.

According to Meta's announcement, Manus has so far served millions of users worldwide. As of early December, since its launch, Manus has processed over 147 trillion Tokens and created over 80 million virtual computers.

Manus's growth miracle is unique to the AI era, which has been accelerating at an unprecedented pace. In less than a year, the three core members of Manus have achieved remarkable feats, which will greatly inspire young AI entrepreneurs to ride the wave.

However, the fact that Manus was acquired by an overseas tech giant has also sparked a sense of regret.

Yu Bo, the founder of YouMind and Yuque, said on social media that he always felt there was a tinge of sadness to the story. One reason for the sadness is that Manus ultimately failed to realize the dream of becoming an independent company. The cost of AI entrepreneurship is extremely high, and it takes wisdom to adapt to the situation. If it hadn't been acquired at this point, the story might have taken a turn for the worse. General Agents are ultimately a game for the giants. The second reason is: Why Meta? Just as Baidu was anxious about securing a mobile foothold and acquired 91 Assistant, Meta is now anxious to break through in AI. We hope history doesn't rhyme and that a different story unfolds.

How did Manus grow from being labeled the "king of shell wrapping" to becoming the "darling" of a large company? On some key issues, although many have criticized Manus for "taking shortcuts," it has made the most pragmatic choices in line with the trends. Looking back at its growth path now, we can better understand the underlying logic of AI entrepreneurship.

01

General Agent vs. Vertical Agent

Although Manus's demonstration in March was stunning, the idea of a general AI Agent still doesn't have the majority support in the industry.

An investor who passed on investing in Manus told China Entrepreneur: They were not optimistic about general-purpose Agents and didn't invest because they couldn't believe that a startup could scale such a business.

He gave two reasons: First, if you define something too generally, you'll always fail to meet 20% or 30% of the long-tail needs. As a general product, failing to address the long-tail means you can't build a good reputation. "Big tech companies might be more suited to this task, as they have the resources, data, and context." Second, developing a general-purpose Agent consumes a large amount of Tokens and incurs high inference costs. Passing these costs on to consumers would make it difficult for them to accept.

In fact, the user experience of Manus's general Agent has been mediocre, which is also the focal point of the debate over whether Meta got its money's worth in this deal.

Many users believe that Manus has limited capabilities in professional vertical scenarios, and there are powerful tools like Claude and GPT for creating PPTs and websites. "Manus is also more expensive." However, some users say that using Manus to do "dirty work" mindlessly is a great experience. "It's very labor-saving to use Manus for tasks like organizing reports and scraping data without having to search for tools and ports one by one. It's also very convenient for conducting research and performing tasks that require a lot of trivial and simple work."

When discussing the debate between vertical and general Agents in an interview, Ji Yichao divided the Agent products on the market into two categories: One type is the "rule-driven" Agent. Developers pre-design the workflow, and the Agent follows a specific branch based on the situation. The advantage is that it's controllable and predictable, but it can only operate within the pre-designed workflow.

The other type is the "intelligence-driven" Agent advocated by Manus. After building the model platform, there's no pre-set workflow, and the model makes its own judgments. This type of Agent can handle unexpected situations flexibly but is also less stable and might act stupidly at times.

On December 27, 2025, Zhang Tao, the product manager of Manus, published an article titled Meeting Minutes of Manus Project Initiation on his official account. He compared the two different development paths of Agents to the model differences between Baidu and the web navigation site Hao123:

Chatbot/Hao123 model: Developers, as the "supply side," pre-implement and integrate various specific functions (links). What users can do is limited to the capabilities provided by the developers. This model has slow expansion and is prone to homogeneous competition.

Agent/Baidu model: First, build a powerful underlying platform with general capabilities (like a search engine that can crawl and understand everything). Due to its generality, this platform attracts a large number of users to try various tasks (Queries). Then, by analyzing high-frequency and high-value Queries, optimize the platform in reverse and introduce "preset capabilities" (Presets) like "box computing" or "Aladdin cards" to enable common tasks to be completed in seconds.

In addition, when a general Agent calls the model, the intersection of underlying capabilities may also lead to the "generalization" of product capabilities. Ji Yichao mentioned in an interview that they added an image recognition ability to Manus, and after that, Manus learned on its own to check whether a website was functioning properly.

Despite these issues, there are still investors who firmly support Manus. The aforementioned investor told China Entrepreneur: They're mainly investing in the team. "Maybe Manus won't succeed, but they believe that the products developed by this team will eventually have one that makes it big."

Among the firm supporters is ZhenFund. Liu Yuan, a partner at the institution, led the investment in the Manus team. After the deal was announced, Liu Yuan wrote an article on ZhenFund's official account, saying that ZhenFund had accompanied Xiao Hong and his team for a full decade, making five consecutive investments. The fund also provided several partners and excellent young employees to the company, including former ZhenFund employees and interns.

The investor who passed on investing in the Manus team told China Entrepreneur, "We're still in the very early stages of this wave. The outcome is still uncertain. We've also invested in many vertical Agent projects, and their commercialization and subsequent financing have been quite successful."

02

Engineering Capabilities vs. Model Capabilities

From a technical perspective, there aren't many secrets in Manus's architecture. It calls on large language models from companies like Anthropic at the underlying level, which has led to accusations that its innovation is limited to product packaging and process orchestration, making it just a "shell-wrapping" product.

However, compared to model capabilities, Manus's competitiveness lies more in its engineering capabilities. Its core value lies in understanding tasks, scheduling tools, and the execution framework. Manus believes that with the evolution of large models, it will be able to achieve complex goals more efficiently under the strength of its engineering capabilities.

Xiao Hong said in a media interview, "Model capabilities are evolving rapidly, but the 'shell' also needs to evolve. After each generation of model capabilities evolves, it's not necessarily the original manufacturer but a third-party vendor that presents the user-perceivable value."

Liu Yuan also mentioned that although the (Manus) team doesn't engage in model training, they were among the first to recognize that a deep understanding of users and excellent product engineering capabilities are highly scarce in this era. Although these capabilities are often not regarded as "core technologies," they are the decisive factors for a company and a product to succeed.

This is also the core of the competitiveness of AI products - taste. As Steve Jobs once said, taste determines everything. According to Zhang Tao's description, the UI (user interface) design philosophy of Manus is as follows:

First, Progressive Disclosure. By default, what should be presented to the user is an extremely simple interface (possibly just a dialog box). As the task unfolds, the tools used by the Agent will "emerge" as independent windows or tabs.

Second, OS-like Metaphor. Design different core functions (such as browsers, spreadsheets, and document editors) as independent and equal "first-level applications" rather than having them nested chaotically.

Ji Yichao summarized the product philosophy as: It's more important to do a thousand small things right than to do three big things right.

For example, the operation of the Agent currently still largely depends on the MCP protocol developed by Anthropic. However, most current Internet services are not connected to MCP. Manus's strategy is to downgrade it, call the API to read documents, or simulate human behavior to scrape data using a browser.

This enables Manus to handle a wider variety of situations and provides a smoother user experience compared to similar products.

03

Domestic vs. Overseas

Manus's high-profile acquisition has also shown startups the opportunities in the overseas market.

Currently, international Chinese startups are dividing into two categories. One group focuses on overseas markets first and then returns to the domestic market. For example, PixVerse under Aish Technology launched its domestic version, "Paiwo AI," in June 2025. The other group goes directly overseas as an independent company, like Manus.

In July 2025, Manus withdrew its headquarters from China and relocated the entire team to Singapore. Although there were many compelling reasons for Manus to go overseas.

An investor told China Entrepreneur that overseas users are more willing to pay for AI products than domestic users. More than 90% of European and American users choose to pay annually, while Chinese users generally opt for shorter-term subscription models.

Another highly regarded AI star entrepreneur in 2025, Chen Mian, launched Lovart, the world's first design Agent, overseas in July 2025 after founding LiblibAI. The number of applicants exceeded 100,000 within five days of its launch. During the product testing period, it was deeply used by nearly a million designers and creative professionals.

Chen Mian's strategy is to enter the market first and then deal with competition. He once told China Entrepreneur: The window period determines whether you can enter at the right time. By leveraging your first-mover advantage to build your own unique barriers, you can truly survive.

After visiting Silicon Valley in May 2025, Xiao Hong also said on social media: Our survival is not a given. Day by day, as a product and a company, we must earn the right to survive through our own value.

Next, for Manus, aligning with an overseas tech giant will also involve a choice of camp. In June 2025, Anthropic cut off the model API interface for the AI Coding startup Windsurf because OpenAI planned to acquire it. Although the acquisition plan later fell through, Windsurf was dismembered, and its core team was acquired by Google for $2.4 billion. Whether Manus will face a similar challenge from Anthropic remains unknown.

This article is from the WeChat official account “China Entrepreneur Magazine” (ID: iceo - com - cn), author: Yan Junwen, editors: He Yifan, Li Yuan. Republished by 36Kr with authorization.