Starring Manus and Jing Kun, a global carnival of Chinese AI Agents.
Text by | Zhou Xinyu
Edited by | Su Jianxun
At the end of June 2025, at a technology summit in San Francisco, USA, a flyer appeared on each seat in multiple closed - door meetings. This A4 - sized paper, with black text on a white background, proudly printed two large lines:
$36 Million ARR in 45 Days, 8 Products in 10 Weeks.
This remarkable achievement comes from a general AI Agent, Genspark. The founder of the company behind it, MainFunc, is Jing Kun, the former vice - president of Baidu Group and the CEO of Xiaodu Technology.
“Overseas enterprises were shocked and kept asking what the company behind Genspark was,” an AI practitioner based in Silicon Valley told “Intelligent Emergence.”
On May 19, 2025, Jing Kun, the founder and CEO of MainFunc, posted Genspark's ARR on X.
Since 2025, the global tech circle has surely set its eyes on some Chinese - origin AI Agents.
The aforementioned practitioner recalled that many Silicon Valley VCs at the exhibition mentioned Genspark and Manus, the first - generation top - tier Agent. “The last time Chinese - origin AI received such attention was when DeepSeek V3 and R1 were released.”
Not only VCs, but also tech KOLs like Elon Musk have left positive comments on Chinese - origin Agents such as Manus and Lovart on social media.
All of this is reasonable. One cannot ignore Manus, the “key” to the Agent wave. This general Agent developed by the Monica team has achieved several miracles globally:
In March 2025, the month when Manus was released, it reached 23 million MAU.
A month later, Manus was reported to have received a $75 million financing led by Benchmark, and its post - investment valuation quickly exceeded the $500 million mark. This is a particularly scarce overseas financing for Chinese teams given the current tense geopolitical situation.
Source: Manus official website
It can be said that in 2025, Chinese - origin teams initiated a global Agent frenzy.
AI Agent refers to an AI application form that can autonomously plan tasks, call external tools, and deliver results. In other words, AI is not just for simple chatting but can do practical things like humans.
This is not a new concept. In June 2023, Lilian Weng, then the head of application research at OpenAI, published a 6,000 - word article, “LLM - powered Autonomous Agents,” which quickly reached a consensus in the emerging AI application layer: to develop Agents.
However, limited by the capabilities of the underlying models, the “brains” providing intelligence, most Agents at that time were still confined to “chatting” or RPA (Robotic Process Automation) frameworks and failed to achieve real - world implementation.
Later, in November 2023, with the release of OpenAI's low - code development tool, GPT Builder, Agents quickly lost momentum amidst the doubts that “OpenAI had lowered the development threshold, leaving no moat for startups.”
An investor who invested in several Agent applications in 2023 frankly told “Intelligent Emergence,” “These applications didn't perform well in 2024. Eventually, everyone started to return to the traditional RPA route and struggled with customization.”
In 2025, there was a “renaissance” of Agents, and entrepreneurs finally caught the real wave.
Different from the “theoretical discussions” in 2023, the Agent wave in 2025 is based on objective technological leaps, a new product form that has become a consensus, and visible user scale and revenue.
Chart by “Intelligent Emergence”
Among them, Claude and Manus, an overseas model and a domestic application respectively, became the fuse for this Agent explosion.
The former not only released the hybrid reasoning model Claude 3.7 Sonnet on February 25, 2025, which significantly improved the programming and development performance of the model, but also released the MCP (Model Context Protocol) in November 2024, which is crucial for the Agent ecosystem. This means that AI can freely call external tools supporting MCP without cumbersome adaptation.
There's no need to elaborate on Manus' popularity. More than marketing strategies, Manus' greatest significance lies in defining the product paradigm of Agents for confused practitioners.
It can be observed that subsequent Agent applications such as Lovart, YouWare, and Fellou 2.0 all bear the shadow of Manus: a dialogue box showing the thinking chain, paired with a visual panel for task execution.
Manus task execution interface. Source: Trial by “Intelligent Emergence”
An insider told us that LiblibAI, the company behind the Agent Lovart, planned to develop design - oriented Agents at the beginning of 2025 but was struggling to find a product form until Manus was released. He mentioned that the Lovart team started aggressive R & D three days after Manus' release.
However, behind the prosperous scene, Agents are anxious about how to maintain growth.
The growth dividend created by novelty is gradually fading. According to SimilarWeb's monitoring data, since 2025, Manus' monthly visits have dropped from 23.76 million in March to 17.3 million in June, and Genspark's monthly visits have decreased from 8.88 million in April to 7.69 million in June.
The unpredictable geopolitical situation and policies also bring uncertainties to Chinese - origin Agents, especially in their overseas business development.
Manus has made its choice. Since June 2025, signs such as laying off Chinese employees, clearing Chinese social media accounts, establishing a Singaporean team, and showing “Not available in your region” on the Chinese website indicate that Manus is exiting the Chinese market.
Manus explained that this was an adjustment for operational efficiency and an international layout. However, more people believe that this move is actually related to the US Treasury Department's supervision of overseas investments.
“Several investors in Manus, including Benchmark, should have faced pressure from the US Treasury Department,” an investor analyzed. “In the future, companies seeking overseas funds will likely face the dilemma of choosing between markets.”
It only took 3 months for the Agent industry to go from a high - profile start to anxiety.
An AI practitioner met Jing Kun, the founder of Genspark, and his team at the VentureBeat Transformer Conference. He recalled to “Intelligent Emergence” that the anxiety of this star entrepreneur was obvious.
He heard Jing Kun ask his employees a question: What should Genspark do next?
ARR Soars to Millions in No Time: General Agents Break Through the Dominance of Big Tech
Looking at this intensifying Agent craze, there are several interesting phenomena:
On the one hand, Agents achieve high revenues at an astonishing speed. For example, Genspark only took 9 days to reach $10 million in ARR. In contrast, Cursor, a star company in the AI Coding field, took 21 months to achieve the same milestone.
Of course, the statistical caliber of ARR may need further verification, but the rapid revenue growth of Agents proves one thing: After two years of market education, the willingness of domestic and foreign users to use AI to solve problems is increasing, which provides a fertile ground for AI entrepreneurs to realize quick monetization.
On the other hand, general - purpose products, which have always been the comfort zone of big tech companies, first emerged and were validated among startups in the Agent track.
“Intelligent Emergence” has counted the layouts of startups and big tech companies in general Agents since 2025. It is obvious that in terms of the number of products launched, the progress of commercialization, and the influence of products, startups are more aggressive.
An insider told “Intelligent Emergence” that Baidu's general Agent, “Xinxiang,” has not been monetized yet, and its DAU is not high. After only 3 months of release, “the company stopped aggressively investing resources in it.”
Chart by “Intelligent Emergence”
This is a phenomenon that goes against the common perception.
Big tech companies, which are good at developing large - scale products, are often restricted by organizational inertia and caution towards policies and regulations in the era of AI innovation.
An employee at Baidu told us that since “Xinxiang” focuses on the domestic market, it cannot freely access high - performance overseas models such as Claude and Gemini. “There are limitations in functionality, and the team has to spend a lot of effort on engineering issues.”
Flexibility and efficiency have become the greatest advantages for startups to challenge big tech companies in the AI era.
@ZengyiQin, a Ph.D. from MIT, commented on Manus on X, which is a portrayal of the rapidly rising general Agents: “It is a good product, but not a technological breakthrough.”
To some extent, Manus' product philosophy has explored a way out for current AI application startups:
Respond quickly to the latest technologies, efficiently build products based on third - party models and other technologies, and conduct growth and commercialization verification in the more mature overseas market.
Startups need to excel in two aspects. First, they need to be fast to establish user awareness.
A 50 - person team only took 3 months to develop Manus. An insider told us that Lovart took less than 2 months from aggressive R & D to launch.
The dividend period of technologies and concepts is limited. It is obvious that only Lovart, which was launched closely following Manus, managed to attract some attention. Agents launched later under the banners of “the world's first xxx agent” or “Manus in the xxx field” generally made little splash.
Another thing entrepreneurs need to work on is to find application scenarios, define product forms, and polish the user experience with necessary engineering means.
Even though Manus “wraps” multiple models, the view of Xiao Hong, the founder of Monica, “the wrapper has its uses”, has become a consensus. In a blog post shared on July 18, 2025, Ji Yichao (Peak), the co - founder and chief scientist of Manus, shared the efforts the team put into context engineering: They reconstructed the agent framework 4 times to improve the running speed and scalability of the Agent.
Is Specialization the King's Way for Entrepreneurship Instead of Generalization?
Many Agent products seem to be making a big splash but are actually not very useful.
For example, many users commented that even when performing basic tasks such as website generation and research report analysis, Manus still has problems such as “hallucinations,” inconsistencies, and poor intention understanding. The actual experience is not as amazing as the official demo.
The decline in data also means that the PMF (Product - Market Fit) of general Agents has not been fully verified. Since their release, the monthly visits of Manus and Genspark have generally shown a downward trend.
“It's better to use Claude 4 directly instead of Manus if you're spending money. The quality of the model's output is higher,” an industry practitioner told “Intelligent Emergence.”
This also points out the disadvantage of startups in the general Agent track: the product performance is overly dependent on the intelligence of third - party models.
“In the future, general Agents will most likely be developed by companies like OpenAI and Anthropic (the company behind Claude) that possess top - tier models,” an AI investor told “Intelligent Emergence.” “For these model providers, developing general Agents is just a piece of cake.”
A typical example is that about 20% of users upgraded to the $200 - per - month Pro membership to use OpenAI's Agent “Deep Research” for ChatGPT Pro users.
As for the future of “Manus - like” products, the aforementioned AI investor made a judgment to “Intelligent Emergence”: “Ultimately, general Agents developed by entrepreneurs will gradually transform into specialized Agents.”
Currently, many AI companies that started in specialized fields have extended their business to specialized Agents. For example, LiblibAI, which originated from an image model community in 2023, launched the design Agent Lovart.
Cai Haoyu, the former founder of Mihoyo, also released his first AI game, “Whispers from the Star,” in which the characters are Agents. Users can trigger different storylines by interacting with them.
Chart by “Intelligent Emergence”
Even though the current model performance cannot support large - scale and comprehensive general tasks, when the application scenarios are narrowed down to specific vertical fields, Agents can already meet many needs.
At the same time, the accumulation of knowledge and resources in vertical fields is the only barrier for current AI applications. This affects the product design of vertical Agents and their integration into workflows. For startup Agents, this is a relatively safe niche in the face of the encirclement of big tech companies.
Looking at this Agent wave, although the situation is still uncertain, no one doubts that Agents will become a short - lived bubble like in 2023 in the face of obvious technological and product changes.
