AI agents are in a hot trend, and investors are both anxious and fearful.

An investor from a RMB fund is busy contacting AI Agent entrepreneurs. Recently, she has been troubled by the fact that many US dollar funds in the industry are "spreading money widely" on Agents. Her own fund is not that large in scale. How can she grab a share in this field?

The focus of capital is shifting rapidly. Last year, this investor was looking at AI hardware, while this year, AI Agents have taken the center stage. After the general AI Agents Manus and Genspark became extremely popular in the spring of this year, investors shifted their attention to the AI Agent field.

This year, the investment market for AI Agents has shown an explosive growth trend. According to the international venture capital database Crunchbase, AI Agents have become the primary trend in seed - round investments in 2025. As of June 2025, AI Agent startups have raised a total of about $700 million. The market research firm MarketsandMarket predicts that the market size of AI Agents will soar from $5.1 billion in 2024 to $47.1 billion in 2030, with a compound annual growth rate of up to 44.8%.

Startups in the AI Agent field have broken the norm in terms of both the speed of financing and the climb of valuation.

Anysphere, the company behind the well - known AI programming Agent Cursor, saw its valuation nearly quadruple in six months, rising from $2.6 billion to $9.9 billion. Lila Sciences, an Agent company in the life sciences field founded this year, received a huge investment of $200 million in its seed round. The valuations of both Manus and Genspark have reached $500 million, which are five times and twice their valuations last year respectively.

In the past two or three years, the input and output of large foundation models have been disproportionate, and coupled with the unclear profit prospects, capital is prompted to look for more economically viable investment directions. The appeal of Agents lies in that, when linked to productivity, they accelerate the commercialization story of AI integration into various industries.

The year 2025 is known as the "Year of Agents". With the emergence of DeepSeek and heavy investments from large companies, large models have become more and more popular. Naturally, applications represented by Agents have become a new direction, and almost all technology companies have deployed Agents this year. While participating in the upsurge, investors are also worried. The valuations that have been rapidly pushed up in the short term also mean higher market and commercialization requirements, but this industry is just in its infancy.

01 Targeting Star Teams and Spreading the Net Widely

In the first half of this year, there were frequent large - scale financing events for Agents.

According to data from the international venture capital database Crunchbase, among the 16 Agent startups in the first half of the year, the seed - round financing amounts ranged from $9 million to $200 million. Among them, there were 10 vertical Agents, accounting for more than half, covering industries such as medicine, construction, logistics, and manufacturing.

In addition to startups, large technology companies also regard Agents as one of the tracks. The four major US technology giants are all building their own Agent ecosystems. Google is building an Agent matrix covering office collaboration, cloud computing development, and vertical industry scenarios based on the Gemini large model; Microsoft is promoting enterprise - level Agents driven by Copilot, Meta is trying to use Agents to promote the evolution of AI assistants for social products, and Amazon is focusing on Agents for supply - chain optimization.

In China, the rapid change in investment confidence in the Agent market is driven by several key events.

In March this year, Manus, claiming to be the "world's first general AI Agent", went viral across the internet. On the day of its release, more than 150 AI Agent concept stocks in the A - share market hit the daily limit, and companies such as Lifang Holdings and Kurtz Group saw gains of more than 20%.

A working area at Manus' Wuhan headquarters on April 18, 2025. Photo/Visual China

In April, Genspark, founded by a former Baidu executive, transformed its AI search into a general AI Agent product. It only took 9 days to exceed $10 million in ARR (Annual Recurring Revenue), and achieved $36 million in ARR 45 days after its launch.

Investors saw hope from both the market response and users' willingness to pay. An investor told Caijing that AI Agent applications are more in line with the essence of business, and the investment market is "competing in speed, momentum, and cognition".

The investor mentioned that compared with RMB funds, US dollar funds are more aggressive, focusing on investing in people and spreading the net widely. According to her observation, several leading US dollar funds had more than 20 Agent investment deals in the first half of the year. They tend to choose leaders with backgrounds from large companies, practical experience, strong technical backgrounds, and product capabilities, or typical young geniuses and scientist - type entrepreneurs.

This trend is driven by the FOMO (Fear of Missing Out) sentiment of US dollar funds. ZhenFund has invested in AI companies such as Manus and Genspark. Its managing partner, Dai Yusen, once mentioned in an interview that ZhenFund has a strong FOMO mentality. "In early - stage investments, the biggest fear is not investing wrongly, but missing out. If you invest wrongly, you may just lose money; but if you miss out, you may miss out on 100 - fold or 1000 - fold returns."

Under the investment upsurge, a new phenomenon is that many Agent projects have received high - amount financing without having products or revenues.

An investor told Caijing that some Agent startups require investors to "continuously invest" in two rounds to get an entry ticket.

"Actually, there is no very rational benchmark to determine how much a person is worth. It mainly depends on market sentiment," said the investor. Since Agents are still in the early stage of development, the valuation method usually involves making horizontal comparisons with the financing levels of other projects in the same track. Once a technical talent with an excellent background plans to start a business, other investment institutions will flock to it, driving up the market price and competing for shares.

02 From Burning Money to Pragmatism

Amid the hype, many investors are still on the sidelines.

The challenge faced by the primary - market financing of Agents currently lies in how to judge a company's moat when the technical barriers are not obvious.

Agent companies mainly rely on the capabilities of underlying large models, and there are high - profile disputes over "shell - like" applications. Since Agent applications are not yet mature, many investors currently evaluate Agent projects mainly based on the actual effects of the products, such as how many complex actions they can complete and what kind of results they can actually deliver. This is related to whether users are willing to "pay the bill" and realize commercial value.

The founder of a vertical - field Agent company mentioned that whether an AI Agent is usable and user - friendly depends on the strength of the underlying large model. Since startups do not have the ability to develop large models, the common practice in the industry currently is to "rely on a big tree". He chose to cooperate deeply with ByteDance's Doubao large model and received investment from ByteDance. However, the valuation given by ByteDance is lower than the company's actual valuation, and "it's hard to tell a more imaginative story, which will also affect the subsequent growth space of the valuation".

Different from the logic of US dollar funds of "investing in people", the aforementioned RMB - fund investor's criterion is to see if the people and the direction match. First, whether the direction is an opportunity for startups. Second, whether investing at this current node will increase the survival probability of the team. "If a certain direction is an opportunity for large companies and startups can't make money in this direction, then the company's valuation is inflated."

On the financing side, some Agent companies also have concerns about the source of funds. They usually focus on the "going - global" strategy and consider whether the funds can support the cash flow and overseas implementation.

The trend of VC (Venture Capital) moving closer to PE (Private Equity) is more obvious. The institution where the aforementioned investor works has always focused on early - stage technology investments and has invested in many embodied - intelligence projects before. Affected by the sentiment in the investment market, this institution often discusses how to break through, but she can't come up with a good solution.

She is looking forward to the emergence of a commercial inflection point for Agents. Currently, she can only contact Agent entrepreneurs and continuously follow up on their growth paths on the one hand; on the other hand, she continues to observe and pay more attention to the Agent companies that remain after the early - stage elimination, and then invest in subsequent rounds.

The financing enthusiasm in the AI field is still declining. According to data from Zero2IPO Research, in the first quarter of 2025, driven by the multiple upsurges of large models, embodied intelligence, and AI Agents, there were a total of 351 financing cases in the domestic artificial - intelligence field, with a total financing amount of more than 20 billion yuan, a year - on - year decrease of 20.5%.

Investors are becoming more pragmatic and care more about data related to commercial monetization.

Zhu Xiaohu, the managing partner of Shunwei Capital, previously said in a media interview that current AI applications can be invested in based on a data - driven logic without burning money. The test standard for a company's B - end applications is to achieve $10 million in annualized revenue within 6 - 12 months, and a good company can achieve it in 6 months. "This kind of logic allows me to sleep well at night. Basically, it's a project where the money invested this month can be recovered next month. It's really amazing," he said.

However, the traditional revenue - prediction logic is also facing challenges. In the past, ARR (Annual Recurring Revenue) was important because SaaS companies mainly obtained revenue by selling software functions or usage rights, which made product revenue highly predictable. But now, the business model of Agents is shifting from "providing tools" to "delivering value", and it tends to charge based on "delivered results". The monthly revenue linked to usage volume is more volatile.

An investor told Caijing that some Agent entrepreneurs use the ARR indicator to mislead investors. Since early - stage revenue may come from test customers, users may make large - scale purchases in the short term and then quickly cancel their subscriptions because the product fails to meet their expectations, resulting in an inflated ARR.

Some investors are also exploring new measurement indicators, focusing on the health of revenue, the repurchase rate, and the sustainability.

Currently, startup financing in the AI Agent field can be roughly divided into three categories: The first category is star teams. Most of them are in the sight of a small number of investment institutions in the early stage and can quickly obtain high - amount financing; the second category is AI companies that have transformed from traditional applications and tools. They have accumulated some customer resources and market experience and are seeking financing during the market upsurge; the third category is those that rely on the ecosystem of large companies, can start quickly, and have the opportunity to be acquired by large companies.

The first - category startups are currently the most popular among investors, but the risks of subsequent financing and development are relatively high. The latter two categories are more stable, but the space for valuation growth is also limited.

03 Racing Against Investment Sentiment

Before investors' concerns grow, entrepreneurs need to achieve results quickly. Time has become the scarcest resource for Agent entrepreneurs.

Tian Yihao, the founder of the e - commerce Agent company Turing Bazaar, has set a deadline for himself: three months. Within three months, "try to make this product a hit". Otherwise, if the window period is missed and the number of imitators increases, the company will fall into a "slow - torture" dilemma.

Quickly occupying users' minds is the common goal of all Agent startup teams. Behind this approach, there is a gradually clear industry practice.

First, seize the opportunity with a "concept". Yue Kun, the founder of the Agent startup company Danheji, is preparing for the public - testing period of an Agent product. He believes that in the early stage, for an Agent company to impress investors, it either needs to have a remarkable market response or present real - world payment data. Similar to Manus' claim of being the "world's first general Agent" back then, his product is promoted as the "world's first Agent glasses".

Second, create buzz through "publicity". Just like the popular invitation - code mechanism in communities, "Build in Public" has almost become a standard for startups, focusing on "marketing while developing". The Agent design company Lovable held a live - streamed competition between traditional designers and its own AI functions in May, which attracted more than 120,000 views, detonated the community, and promoted secondary dissemination.

Then, leverage capital with "data". Tian Yihao's plan is to quickly verify core data such as customer - acquisition cost and user - retention rate within about a month after the product is launched. Once the data model is proven to work, this "report card" will be the best tool to impress the next - round investors, thus enabling the business to develop continuously.

Startups also need to face the problem of large companies entering the race. An investor from a RMB fund told Caijing that in the current situation where technology and data are becoming increasingly transparent, the only barrier for investing in an Agent company is to "run fast". The bet is that it can use a faster iteration speed and a deeper industry understanding to build its own user perception and brand barrier before being covered by giants.

Considering the user trust in the product, Tian Yihao introduced well - known brand customers to use his product. His company's valuation reached 20 million yuan only one month after its establishment. When seeking investment, his first choice is not traditional investment institutions but "customers". He believes that investment from customers is the most convincing, as it can not only quickly "endorse" the product but also provide certain resource support.

Even if the startup teams have a fast iteration speed, in a volatile market, it is still uncertain whether the first - mover advantage can be sustained. Investors represented by Zhu Xiaohu believe that general Agents will ultimately be an opportunity for large companies, while niche Agents in vertical fields are the breakthrough points for startup teams. Other investors believe that the definitions of general and vertical are only for the early stage, and the tracks will transform flexibly in the future. They still need to wait and see who will survive this elimination round.

This article is from the WeChat official account "Caijing Magazine", author: Huang Siyun, editor: Liu Yiqin. Republished by 36Kr with permission.