Executives from large companies and prodigy youths are flocking to Agent startups.

AI Agents are becoming the "work partners" for an increasing number of people.

At the Baidu Developers Conference in May this year, Robin Li proposed that in the AI era, the measure should not be how many tokens are consumed, but DAA (Daily Active Agents), which refers to how many agents are delivering results to humans every day.

Behind this is a competition among major tech companies to launch Agent platforms for ordinary users: ByteDance's Coze, Baidu's AgentBuilder, Tencent's Yuanqi, and Alibaba's Bailian. Almost every base model provider is promoting its own Agent development platform, hoping to transform the way people work.

Another group of people is delving deeper into AI applications - starting AI Agent startups. Since 2025, AI Agents have taken over from generative AI large models and become the most crowded track in the startup investment market. It's easy to tell stories to investors, but building a profitable business requires breaking through numerous "seemingly beautiful" illusions.

For this reason, "Bao Bian" interviewed several AI Agent entrepreneurs to explore the current situation, opportunities, and challenges of Agent startups.

Another Startup Wave Boosted by AI

The wave of AI Agent startups began in early 2025. Manus was the first to enter the market, igniting investors' imagination about workflow transformation. Subsequently, similar products from major tech companies quickly followed, and AI Agents became a popular track for seed - round investments in 2025.

Wang Yuxuan, a student at a science and engineering university in Beijing, also tried his hand at entrepreneurship during this period. He designed a Prompt workflow to improve the effect of AI image editing and has been looking for suitable AI Agent startup projects ever since. Almost at the same time, another entrepreneur, Wei Longjie, founded Ayu Law, a legal data compliance AI Agent.

The boom in Agent startups is the result of multiple factors.

The AI programming ability has significantly lowered the threshold for "product creation". Tools like Cursor, Lovable, and Claude Code, representing "Vibe Coding", enable non - professional developers to quickly build prototypes. "Making a product" has become extremely simple.

Internet giants have also contributed talents to the startup army. Jing Kun, the former vice - president of Baidu, founded MainFunc after leaving the company and launched the AI Agent product Genspark. It received $60 million in seed - round funding and completed three rounds of financing totaling over $400 million within a year and a half, with a valuation of $1.25 billion. Wang Ming, the former vice - president of DingTalk, founded Panfeng Intelligence in October 2025 and has received tens of millions in financing so far. Its content e - commerce Agent OS Moras focuses on automated product selection, script generation, and data analysis.

As of early 2026, more than 30 AI companies have been founded by former ByteDance employees. Lin Junyang, the technical leader of Alibaba's Tongyi Qianwen, and several core members of ByteDance's Seed have also recently joined the startup army.

These entrepreneurs, with an understanding of platform effects and traffic operations, are trying to replicate the growth myths of the early Internet era. In early 2026, Zhu Fei, who used to work at Meituan, planned Quote.law, an AI Agent collaboration platform for legal professionals, with two other founders.

AI entrepreneurs have their own industry salons, which are the "hackathons" that have sprung up across the country. A hackathon is a combination of "hack" and "marathon". It is a collective programming activity that originated in Silicon Valley. Participants are required to form teams freely and complete a demonstrable software or hardware prototype from scratch within a 24 - to 72 - hour closed period. The final evaluation is conducted on - site by judges.

In the past two years, first - tier, second - tier, and third - tier cities, top universities, and technology giants have all organized various AI - themed hackathons. Wei Longjie participated in a recent hackathon in Nanjing, hoping to gain some attention for his project and find some collaborators.

Relevant data shows that the market size of the Chinese AI Agent industry reached 18.234 billion yuan in 2025, a year - on - year increase of 78.03%. The industry has entered a period of explosive growth. The 2026 government work report also included "agents" for the first time, indicating an upgrading of the national strategic positioning.

The AI Agent financing field is quite hot. Leading projects have strong capital - attracting capabilities, and their valuations are rising. For the attempts of a large number of small and medium - sized entrepreneurs, many investment institutions prefer an investment strategy of "less but more diversified".

From conversations with some entrepreneurs, it can be seen that venture capital is becoming more cautious and tends to follow leading institutions in co - investing.

The enthusiasm in the secondary market is more obvious. Newcomers like Zhipu and Minimax, which have recently listed on the stock market, have brought multiple - fold returns to investors. Now, companies like Dark Side of the Moon are also queuing up for listing.

Under the technology - driven startup wave, new trends are emerging, such as the youth of core figures.

Wang Yuxuan has been organizing various startup salons recently and has noticed that investment institutions prefer "child prodigies". The so - called "child prodigies" generally refer to students who have achieved computer research results at a very young age, many of whom are even under 18.

The most famous among them is Chen Guangyu, a senior high school student at an international school in Shenzhen. In November 2025, he participated in the R & D of KIMI's large model as an intern. In March 2026, the paper "Attention Residuals", in which he was a co - first author, was published, which received public praise from Elon Musk on social media. In the venture capital circle, algorithm geniuses like Chen Guangyu are a golden ticket for getting financing.

How Can Agents Surpass the Capabilities of Base Models?

Agents rely on the capabilities of base models. So where are the opportunities? The answer is in areas where base models fall short, such as industry expertise.

AI Agents are also penetrating from general applications into specific industries, including the legal industry. "The legal industry is an old and slow - moving industry. Many materials have not been digitized, are scattered, and have complex contexts. A large amount of work still involves inefficient document circulation and repetitive communication. AI provides an opportunity to transform this industry," Zhu Fei said about the role of Quote.law.

In Quote.law, users can organize materials, conduct legal research, edit documents around the same project, and collaborate with AI Agents in the same environment to advance tasks. In a collaborative scenario, writing legal documents by oneself is too troublesome, and using others' documents may cause concerns. AI Agents are a good "third - party".

Quote's long - term vision is to become the "Alipay of the legal field", providing users with credit endorsement and legal services through an AI platform.

Wei Longjie's Ayu Law specializes in B - side data compliance. As a senior lawyer graduated from Peking University Law School, he has refined his years of experience into a high - quality Prompt and Memory database, enabling AI to recognize data regulatory differences in different jurisdictions (such as the US, China, and Europe) and predict the compliance risks that enterprises may face as they grow.

Since his target group is mainly B - side, he usually reaches out to startups in batches through VC institutions and startup communities. He finds customers while looking for investment and cooperation. For some small and medium - sized enterprises that rely on data, the existence of AI Agents is undoubtedly a blessing. The original six - figure legal fees have now been reduced to 20,000 - 25,000 yuan.

There is also the transformation of traditional manufacturing.

"Localized customization of B2B Agents, which penetrates into the enterprise's process management and solves the automation problems of certain business nodes," is how Wang Yuxuan understands "manufacturing - specific Agents".

Among them, there is a company called Yuhe Technology, which completed its angel - round financing in 2024. Its core business is to build an Agent system of "base + private data" for manufacturing enterprises. Taking a shipyard as an example, in the past, it took senior engineers several weeks to complete a pre - sales plan (including ship design, part selection, and quotation generation). Now, by feeding the enterprise's decades - old historical plan data to the Agent, new salespeople can quickly produce professional plans with the help of the Agent.

The common feature of these industry - specific Agents, such as those in the legal and manufacturing process transformation fields, is that base models cannot directly solve professional - scenario problems. The products need to have industry knowledge, the ability to discover and solve problems, and initiative.

Compared with base models, Agents can also improve AI memory and reduce hallucinations through a restricted environment. Zhu Fei regards industry - specific Agents as "sewage treatment plants", which improve the output quality through professional corpus processing and memory optimization.

However, being in a promising track is just the first step. Among the tens of thousands of competitors in the market, what kind of AI Agent startups are more likely to succeed?

Wei Longjie believes that the most important thing is to find real needs. "Some projects seen at hackathons have novel ideas, but it's not clear if they have a market. Users may not be willing to use or pay for them."

Regarding how to verify, Wang Yuxuan believes that "you can first see if you can find 100 users willing to pay. If not, you should change your approach." His previous attempt at AI image editing was just a small trial. Now he wants to work in a venture capital firm or a large company to learn some methodologies. "Once you really start a business, you'll realize that PMF (Product - Market Fit, proposed by Silicon Valley venture capitalist Marc Andreessen and commonly used in strategic analysis of Internet giants and investments) is crucial."

The other side of PMF is that the product can meet the needs, which tests the team's ability. Wei Longjie has rich business experience, and his co - founder is currently pursuing a Ph.D. in computer science in Germany. "When looking for financing, our combination of business knowledge and technical expertise is more likely to attract investors' attention," Wei Longjie said.

Ultimately, there need to be enough users to generate a platform effect. In the Agent track, the platform effect is first manifested as a 'data flywheel': the more users use the Agent, the richer the private corpus, behavior preferences, and industry memory the Agent accumulates, and the higher the quality of the model output, which in turn increases user stickiness. On the other hand, once an enterprise customer embeds the Agent into its business process, the replacement cost is very high.

However, the paradox is that only by crossing the 'cold start' can the flywheel be triggered, and most Agent startup companies go out of business before reaching that stage.

The 'AI Illusions' and 'Human Issues' in Entrepreneurship

Has entrepreneurship become easier?

It seems so. We can now write code with the help of AI and complete products at a very low cost. But this creates new "illusions": thinking that "the hardest part of entrepreneurship is making the product".

In fact, AI entrepreneurship is more brutal. Yupp, an AI evaluation company, received $33 million in seed - round funding but shut down its product within a year of launch. Robin AI, an AI contract tool, received investments from Google and SoftBank but went from its peak to being put up for sale within six months. Humane AI Pin, an AI wearable device, raised over $200 million in total but was finally acquired for only $116 million.

Wang Yuxuan believes that many entrepreneurs are "holding a hammer looking for a nail": "The underlying logic of entrepreneurship has never changed. You can now make products quickly, but the hardest part is always to discover real needs."

On the other hand, market bubbles make many orders not based on real needs but on the pursuit of the AI trend, which is essentially another manifestation of the FOMO mentality. Such needs are not sustainable. Once the technological hype fades or the budget tightens, users will leave.

Moreover, Agent entrepreneurs also face pressure from the 'foundation'.

The capabilities of base models are constantly increasing, which is, to some extent, squeezing the survival space of Agents, such as in terms of memory. The latest versions of mainstream models like GPT and DeepSeek have extended the context to one million tokens. Problems that originally needed to be solved by the Agent architecture are gradually being covered by the native capabilities of base models.

Many Agent startup companies in the market that "only fine - tune prompt words" are facing such survival pressure. Li Kaifu also recently said, "Don't stand in front of the advancing large models. If you do, you will surely be crushed."

Every time the base model takes a leap forward, Agents that rely on prompt engineering and lightweight encapsulation face a "challenge".

In addition, offline data is both an advantage and a problem for industry - specific Agents because it may "never be online". Sometimes, senior practitioners are reluctant to digitize some historical data for self - protection. In more cases, non - standardized practitioner experience cannot be digitized.

This is common in the manufacturing industry. As the old saying goes, "The most valuable assets in a factory are in the minds of old masters." In the medical field, high - quality data is scattered among different hospitals and is difficult to share due to ethical reasons. The same is true in the network security industry, where data itself is the core competitive barrier.

However, compared with business problems, the biggest challenge in entrepreneurship still lies in 'people'.

Zhu Fei believes that finding truly suitable collaborators is often more difficult than raising funds. Ideal candidates should understand complex business and be sensitive to new technologies. Such people usually either hold important positions in mature platforms or have already started independent exploration, and are not easy to recruit through regular channels. Wei Longjie thinks the biggest challenge is communicating with his co - founder. "When collaborating with someone with a pure technical background, you need to explain more things and understand things more deliberately."

In this sense, AI has not lowered the threshold of entrepreneurship. Although writing code has become easier, the threshold of "understanding business, technology, and people" still exists. There has always been such a cycle in the history of technology: first, there is tool worship, thinking that a new tool can solve all problems; then, the bubble bursts, and it is found that the problems still exist.

Only those who can solve the "eternal problems" can stand up when the tide recedes.

This article is from the WeChat official account "Bao Bian" (ID: baobiannews), written by Zhang Jingwei, and is published by 36Kr with authorization.