Topmanager großer Unternehmen und geniale Jugendliche drängen sich in die Agenturgründung.

The AI agent is becoming more and more of a "work companion" for people.

At the Baidu Developer Conference in May last year, Robin Li pointed out that in the era of artificial intelligence, the measure should not be the number of tokens consumed, but DAA (Daily Active Agents), that is, how many agents deliver results for people on a daily basis.

Behind this concept lies a competition among major tech companies for agent platforms for the general public: ByteDance's Coze, Baidu's AgentBuilder, Tencent's Yuanqi, and Alibaba's Bailian. Almost every provider of basic models is promoting its own agent development platform in the hope of changing the way people work.

Another group of people is more deeply involved in AI applications - they are founding AI agent startups. Since 2025, the AI agent has taken over from large generative AI models and has become the most crowded industry in the startup and investment market. It's easy to tell a story to investors, but building a profitable business requires overcoming numerous "seemingly beautiful" illusions.

For this reason, the magazine "Baobian" interviewed several AI agent entrepreneurs to explore the current situation, opportunities, and challenges of agent startups.

Another wave of AI - supported startups

The wave of AI agent startups began in early 2025 when Manus entered the market and sparked investors' imagination about the transformation of work processes. Subsequently, similar products from major tech companies quickly followed, and the AI agent became a popular industry for seed - round investments in 2025.

Wang Yuxuan, a student at a technical university in Beijing, also tried to start a startup during this period. He developed a prompt workflow to improve the quality of AI - based image processing and has since been looking for a suitable AI agent startup project. Around the same time, entrepreneur Wei Longjie founded the AI agent for legal data compliance: Ayu Law.

The enthusiasm for agent startups is the result of the overlap of several factors.

The programming ability of AI has drastically lowered the threshold for "product creation". Tools like Cursor, Lovable, and Claude Code, known for "Vibe Coding", enable even non - professionals to quickly create prototypes. Thus, "creating a product" has become very easy.

Major Internet companies have also brought talents into the startup scene. Jing Kun, the former vice - president of Baidu, founded MainFunc after leaving and launched the AI agent product Genspark. He received $60 million in the seed round and completed three financing rounds within a year and a half, with a total amount of over $400 million. The company's value is $1.25 billion. Wang Ming, the former vice - president of DingTalk, founded Peak Intelligence in October 2025 and has received several million dollars in financing so far. His Content E - Commerce Agent OS Moras focuses on automated product selection, script generation, and data analysis.

By early 2026, former employees of ByteDance alone had founded more than 30 AI companies. Lin Junyang, the technology leader of Alibaba Tongyi Qianwen, and several core members of ByteDance Seed have also recently entered the startup scene.

These entrepreneurs, who understand the concept of platform effects and traffic management, are trying to repeat the growth legends of the early Internet era. In early 2026, Zhu Fei, who used to work at Meituan, and two other founders planned Quote.law, an AI agent collaboration platform for lawyers.

AI entrepreneurs have their own industry salons, so - called "Hackathons", which are springing up like mushrooms. A Hackathon is a combination of "Hack" and "Marathon". Originating in Silicon Valley, it is a collective programming activity where teams have to develop a software or hardware prototype from scratch within a closed period of 24 to 72 hours. At the end, judges evaluate the results on - site.

In the past two years, cities of all sizes, top - tier schools, and tech companies have organized various AI - themed Hackathons. Wei Longjie recently participated in a Hackathon in Nanjing to promote his project and find potential partners.

Statistics show that the market volume of the Chinese AI agent industry in 2025 was 18.234 billion yuan, corresponding to an annual growth rate of 78.03%. The industry has entered a phase of explosive growth. In the government program of 2026, the term "Intelligent Agent" was mentioned for the first time, indicating the increasing strategic importance of the industry.

The financing environment for AI agents is very hot. Top - tier projects attract large amounts of capital, and their company values are constantly rising. However, for many small and medium - sized startups, many investment companies prefer an investment strategy that focuses on "few but diversified" investments.

From conversations with some entrepreneurs, it can be seen that venture capitalists are becoming more cautious and prefer to follow the lead of top - tier investors.

The sentiment in the secondary market is even more obvious. Newcomers like Zhipu and Minimax, which have recently gone public, have rewarded investors with multiple returns. Now, companies like Dark Side of the Moon are also in line for an IPO.

Under the technology - driven startup wave, new trends are also emerging, such as the youthfulness of core personnel.

Wang Yuxuan, who recently organized various startup salons, noticed that investment companies prefer "geniuses". By "geniuses", we mean students who have achieved scientific results in the field of computer science at a young age, and many of them are even under 18.

The most well - known of them is Chen Guangyu, a 12th - grade student at an international school in Shenzhen. In November 2025, he participated as an intern in the development of the large AI model KIMI. In March 2026, he co - published the article "Attention Residuals", which was publicly praised by Elon Musk on social media. In the startup scene, an algorithm genius like Chen Guangyu is a golden asset for financing.

What is the comparative advantage of an agent over basic models?

Agents rely on the capabilities of basic models. So where are the opportunities? The answer is: In things that basic models cannot do, such as professional knowledge in specific industries.

AI agents are also increasingly penetrating from the general public into specific industries, including the legal industry. "The legal industry is an old and slow - moving industry. Much information is not digitized, materials are scattered, the context is complex, and a lot of work still consists of inefficient document flows and repeated communications. AI offers the possibility to transform this industry." This is how Zhu Fei views the role of Quote.law.

In Quote.law, users can organize materials for the same project, conduct legal research, edit documents, and collaborate with AI agents in the same environment to drive tasks forward. In cooperation scenarios, it is cumbersome to write one's own legal documents, and there are concerns about using others' documents. An AI agent is a good "third party" in such cases.

Quote's long - term goal is to become "the Alipay of the legal industry" and offer users creditworthiness and legal services through an AI platform.

Wei Longjie's Ayu Law focuses on data compliance in the B2B sector. As an experienced lawyer who studied at the School of Law of Peking University, he has translated his many - year experiences into high - quality prompt and memory databases. This enables AI to recognize differences in data regulations in different legal jurisdictions (e.g., China, the United States, and Europe) and predict the increasing compliance risks of companies during growth.

Since his offering is mainly targeted at B2B customers, he usually reaches startups through VC companies and startup communities. When looking for investors and partners, he also finds customers at the same time. For small and medium - sized enterprises that rely on data, the existence of AI agents is undoubtedly a blessing. The previous six - figure legal fees have now been reduced to 20,000 to 25,000 yuan.

In addition, there is also the transformation of the traditional manufacturing industry.

"The local adaptation of B2B agents and their integration into corporate processes can improve the automation of certain business processes." This is how Wang Yuxuan understands "agents for the manufacturing industry".

One such company is Yuhe Technology. After the angel financing round in 2024, the company focuses on creating agent systems with "basic model + private data" for manufacturing enterprises. Take a shipyard as an example: In the past, experienced engineers needed weeks to create a sales offer (including ship design, component selection, and offer creation). Today, new salespeople can quickly create professional offers with the support of an agent by transferring the company's historical data over decades to the agent.

The common characteristic of these vertical agents in the legal and manufacturing industries is that basic models cannot directly solve specific industry problems. The products must have industry knowledge, the ability to recognize and solve problems, and initiative.

Compared to basic models, an agent can also improve AI memory through a defined environment and reduce illusions. Zhu Fei sees vertical agents as "sewage treatment plants" that improve the quality of output through professional language material processing and memory optimization.

However, being in a promising industry is only the first step. Among the thousands of competitors in the market, which AI agent startups have the best chance of success?

Wei Longjie believes that the most important thing is to identify real needs. "In some projects I've seen at Hackathons, the ideas were new and interesting, but it was unclear if there was a market for them. Users may not be willing to use or pay for the product."

Regarding validation, Wang Yuxuan believes: "One can first try to find 100 users who are willing to pay. If that's not possible, one should take a different approach." His previous work on AI - based image processing was a kind of test. Now he wants to join a venture capital company or a large tech firm to learn methods. "When actually starting a startup, one will quickly realize that PMF (Product - Market Fit, a term coined by Silicon Valley venture capitalist Marc Andreessen and often used in the strategic analysis of large Internet companies and investments) is of crucial importance."

The other side of PMF is that the product can meet the needs, which tests the capabilities of the team. Wei Longjie has rich operating experience, and his co - founder is currently studying computer science in Germany and working on his doctorate. "When we ask for financing, as a team that understands both the business idea and the technology, we stand out more." This is how Wei Longjie describes their situation.

Finally, there must be enough users to achieve a platform effect. In the agent industry, the platform effect first manifests as a 'data flywheel': The more users use the product, the richer the private language materials, behavioral preferences, and industry knowledge of the agent become. This improves the quality of the model output, which in turn increases user engagement. On the other hand, it is very costly for companies that have integrated their business processes into an agent to replace it.

The paradox is, however, that the flywheel only starts spinning after the 'cold start', and most agent startups fail before reaching this point.

The "AI illusions" and "human problems" in the startup field

Has it become easier to start a startup?

Seemingly yes. We can now write code with the help of AI and develop products at very low costs. However, this can lead to new "illusions": People think that "the hardest part of starting a startup is creating the product".

In fact, AI startups are even harder. The AI evaluation company Yupp received $33 million in the seed round but shut down its product within a year. The AI contract tool Robin AI received investments from Google and SoftBank but went from the top to the sales phase within six months. The AI wearable device Humane AI Pin raised over $200 million in capital but was finally acquired for only $116 million.

Wang Yuxuan believes that many entrepreneurs are "looking for nails with a hammer": "The underlying logic of starting a startup has never changed. One can now develop a product quickly, but the hardest part is still to identify real needs."

On the other hand, market bubbles cause many orders to arise not from real needs but only from the pursuit of AI trends. In essence, this is another form of the FOMO syndrome. Such needs are not sustainable. Once the technological enthusiasm fades or the budget is cut, companies lose their customers.

In addition, agent startups are also under pressure from the 'foundation'.

The capabilities of basic models are constantly improving and are to some extent squeezing the living space of agents, such as memory ability. The latest versions of mainstream models like GPT and DeepSeek have expanded the context to one million tokens. The problems that originally had to be solved by the agent architecture are increasingly being covered by the native capabilities of basic models.

Many agent startups in the market that "only fine - tune the prompt words" are under such survival pressure. Li Kaifu recently also said: "Don't stand in the way of the progress of large models, or you will be crushed."

With each leap in the development of basic models, agents that rely on prompt engineering and light packaging... (The sentence seems incomplete in the original text)