Behind the soaring financing, the investment logic of the Agent track is being reconstructed.

In 2025, AI Agents became the most competitive area in the artificial intelligence industry. On July 17th, OpenAI launched its first AI Agent, ChatGPT Agent, which somewhat resembled the general-purpose agents Manus and GensPark. For the first time, all general-purpose AI Agent startups directly faced the impact from the giant model layer.

ChatGPT Agent integrates the features of two of OpenAI's own Agents. It adds the ability of in-depth research and thinking on the basis of Operator, and adds execution ability on the basis of Deep Research. It can help users handle some complex and implementable tasks. For example, ChatGPT can handle requests such as "Check my calendar and briefly report on upcoming client meetings based on the latest developments" or "Analyze three competitors and create a slide presentation".

Just a few days before OpenAI launched ChatGPT Agent, the code programming tool Windsurf broke off negotiations with OpenAI. Google acquired the core team of Windsurf for $2.4 billion, while Cognition AI, the company behind another programming Agent, Devin, will acquire the remaining Windsurf team. The high cost of talent acquisition and the "split in two" merger and acquisition method became one of the hottest topics in Silicon Valley.

In addition, startups and giants in the entire programming field are also in fierce competition. In June this year, Anysphere, the parent company of the programming star Cursor, completed a $900 million financing round, with a valuation approaching $10 billion. Meanwhile, the large model company Anthropic launched Claude Code, directly entering Cursor's stronghold. Grok 4 was also released in mid-July, with a significant improvement in code capabilities. Google also launched Gemini CLI, directly competing in the programming field.

Moreover, Glean, an Agent in the enterprise search direction, received $150 million in financing, with a valuation of $7.2 billion. HarveyAI in the legal intelligence direction received $300 million in financing, with a valuation of $5 billion. The vertical Agent track is also booming across the board.

In the field of AI Agents, whether it's general-purpose Agents or programming Agents, large model giants are pouring in to seize the living space of startups. As the capabilities of foundation models continue to strengthen, can the core barriers of vertical track Agents withstand the enhanced capabilities of foundation models? How will investors view whether the current AI Agent track is still worth investing in? In this episode of "Silicon Valley 101", host Hong Jun invited Zhang Lu, the managing partner of Fusion Fund, and Zhou Wei, the managing partner of Cyber Creation Ventures, to analyze their investment logic.

The following is a selection of the conversation:

01 The AI Agent financing boom, with star companies emerging in droves

Hong Jun: Which are the leading Agent companies in your knowledge?

Zhang Lu: Personally, I define an AI Agent as a product that can handle complex tasks, not single-line tasks, and can make autonomous decisions.

It can handle a relatively complex multi-step task from start to finish, end-to-end. Of course, the tools it calls during the process still require the team creating the Agent to provide it with a tool library. If it's a vertical Agent, you need to provide it with a knowledge base. This is our basic definition of an Agent, and we use this standard to evaluate which companies we are more optimistic about investing in. For example, when Devin from CognitionAI emerged, it attracted a lot of attention. It is more of a general Agent. There is also Rabbit OS, which is more of an Agent OS (Operating System). A company we invested in, You.com, is also one of Glean's competitors. It recently became a unicorn and is also a very good vertical Agent, and it covers multiple verticals. Several companies have also emerged in the medical field. Another company called Clarity is also doing quite well among vertical Agents recently. In the code field, such as Cursor and Windsurf, if you ask me which one is the best, I actually like Claude Code from Anthropic the most. It's really, really useful.

Source: VCG/Getty Images

A new player, Grok 4, entered the market in the United States on July 9th. I think the timing of our discussion about AI Agents today is very opportune. After the release of Grok 4, our entire team has been using it, and we are very impressed by many of its parameter benchmarks. Of course, it didn't officially release its coding model this time, but from my internal understanding, for example, at xAI, they now write about 70 - 80% of their code using their own internal code model, and the quality is very high. Based on the current performance of Grok 4 in all aspects, I'm really looking forward to its code model. I also want to see how its application scenarios and performance compare with my favorite Claude Code after its release.

Actually, I've talked to many engineers and developers. For example, recently I talked to a group of researchers, and I found that 70 - 80% of them are now using Claude Code. Because Claude Code itself has a very user-friendly programming environment and is very easy to use. Moreover, its complexity and the degree of automatic code programming are more useful than the others. So you'll find that in this field, even though companies like Cursor and Windsurf you mentioned already have so much revenue and financing, the iteration speed is still very fast after new programs and models emerge.

Hong Jun: When we discuss Agents, I find that basically the Agents we discuss are in different fields. Based on my superficial understanding, I roughly divide AI Agents into three categories.

The first category is AI coding, which is what we just talked about with large amounts of financing. For example, Cursor, including Windsurf whose core team was acquired by Google, and Claude Code which you just mentioned. The second major category is general-purpose AI Agents, such as the highly discussed Manus. There are also many vertical Agents mentioned earlier, such as Cresta AI and Harvey AI. Including many medical field companies you invested in, Lu. I think next we can divide them into these three categories and discuss from the perspective of an investor how to view these three categories of Agents.

Source: Cursor

First, regarding AI coding, the first question that comes to mind is that the update of Claude Code from Anthropic has impacted Cursor. So, what is the moat of a programming AI Agent like Cursor? Will it be impacted by large models?

Zhang Lu: I think the evolution of models does have an impact on these AI programmable tools. Of course, Cursor actually has GPT 4 and Claude built-in as code programming assistants. Claude is indeed one of the engines behind Cursor, and they don't only use this one engine. However, Claude Code is a product that can be used directly.

On the developer level, I think Claude Code can be regarded as a direct competitor to Cursor, and the number of migrations is quite large. I've talked to some large-scale technology companies, such as unicorn-level ones. They may have used Windsurf or Cursor more in the past, but now many of them are starting to switch to using Claude Code. So, on one hand, there is a boosting effect. If the models behind you are getting better, it will make the programming companies and AI code companies that have these models built-in perform better.

However, the programming products they develop themselves, because they have their own engines and products, actually have a higher degree of integration. This is why we were quite surprised that after the release of Claude Code, its code capabilities are so strong, especially in supporting complex code analysis, interpretation, and generation, and it is very good at long context analysis. You'll find that whenever something new comes out, we always want to test what level it can reach. Can it handle 100,000 tokens (the smallest semantic unit in text), 200,000 tokens? Can it handle large projects? Claude Code can do it. Also, going back to its developer environment, at least it makes developers in Silicon Valley think it's very useful. Its integration with many plugins, such as Slack and Notion, which are commonly used, is also very good. So, I think Claude Code is not an independent IDE (Integrated Development Environment). In fact, it is more of an enhancement of the code features of the Claude model.

This is why I mentioned earlier that Grok 4 has been using its own internal code model for a long time. If it is now released for third-party use, its performance may be better than other code tools that call or have the Grok 4 model built-in. Of course, on the whole, this market cannot be monopolized by one company. On one hand, the iteration process is very fast. But if a company has already integrated deeply into one programming model, there will be some migration costs. For example, a team from one of my companies really likes Windsurf, and now they don't necessarily have to switch to a new platform because their integration with Windsurf has been quite good.

I also want to mention another new player, Gemini CLI, which is also part of Google's Gemini project and an AI programming assistant. It is more of a terminal for natural language interaction, so it is not a direct competitor. However, it is an important player entering the AI programming market and provides a very good new platform for many companies that want to innovate in this field.

Hong Jun: How did AI coding companies like Cursor, Windsurf, and Devin grow in an environment where Copilot already dominates the market? What were the core breakthroughs and changes at the key nodes of the growth of these large programming software?

Zhang Lu: Of course, we don't invest in the C-end, nor do we invest in any AI code models. But I think one of the very useful aspects of Cursor is its UI (User Interface) and UX (User Experience) design. It is optimized for AI programming and is very well done. If you compare it with ordinary VS Code (Visual Studio Code, a lightweight code editor), its integration process with AI is very smooth and easy to use. And some of its small features, such as auto-completion, automatic code explanation, modification, and context linkage through natural language, are very useful. It greatly lowers the threshold.

I took a programming course before. I don't have a computer background; I have a background in materials science and engineering. But even so, I can use it to build products. Combining it with ChatGPT or Gemini, I can make many interesting small application attempts. So, I think this is one of the very useful aspects of Cursor.

As for Windsurf, I haven't used it much, but several friends of mine like it very much. They think it can better handle multi-step tasks and the process is relatively smooth. It may be a more suitable scenario for building an AI Agent.

Hong Jun: I once read an interview with the founder of Cursor. The interview was actually answering this question: In an environment where Copilot already dominates the code market, how did they break through? I think the rise of startups may not be linear. Instead, at a key time point, when the underlying large model is updated, if the product iteration of the startup keeps up, it can improve the overall user experience. Overall, the feeling from that interview is that it's extremely competitive.

Zhang Lu: It is extremely competitive. One of the core points is that you need to have very strong execution ability, quickly enter the commercial market and achieve rapid revenue growth. And you'll find that the current revenue growth rate is different from the past. In the past, it might take 2 - 3 years for a company's revenue to grow from zero to one or two million dollars. Last year, 70 - 80% of the companies we invested in had their revenue grow 20 times in one year. Some fast-growing companies could grow from zero to tens of millions of dollars, and there was even one company that grew from $500,000 to over $100 million. But the reason for such rapid growth is also because the competition pressure is very high. If you don't quickly enter the commercial market, it will be difficult to maintain your moat. Even if you have high revenue, you still need to keep moving forward. It's really a market of rapid iteration and fierce competition. It's both challenging and exciting.

02 ToB Vertical Agents: A Hidden Battlefield with Low Costs and High Barriers

Hong Jun: Do you think AI coding belongs to ToC (To Consumer, targeting individual consumers) or ToB (To Business, targeting corporate customers)?

Zhang Lu: We also define it as being more towards the ToC end.

Hong Jun: Actually, I'm more interested in the judgment logic of an investor. Everyone is optimistic about the Agent track in the ToB direction, rather than companies in the ToC direction. In which quadrant do they fall in your judgment? Why aren't they on your investment list? Or can they be invested in? Is it worth investing in?

Zhang Lu: In the long term, we are more optimistic about the overall ToB opportunities. So when the AI Agent track emerged, we mainly focused on vertical Agents. We paid less attention to companies in the C-end that are more towards general Agents.

When defining ToB, a more direct scenario is to look at its business model. Is it directly sold to enterprise-level users, small, medium, or large enterprises? Also, we will look at the model itself. Is it a way of directly using a large model as a shell, or can it also create its own vertical models or fine-tuned small models? I think the current competition in artificial intelligence is no longer just a competition between models. More importantly, it is a competition between data and cost. At this time, we need to see which is lower, the cost of directly embedding and using a large model, or the cost of creating a vertical small model for some vertical applications. We are more optimistic about the direction of these vertical small models and edge small models, and the main application scenarios for this direction are ToB, because B-end customers may have higher cost requirements for artificial intelligence products.

Previously, people thought that when it came to verticals, it meant a small market. But now you'll find that many verticals are actually quite large markets. I'll give you an example of a market that you might not usually think of. We invested in a company that focuses on a very niche vertical called commercial paper issuing. It is a short-term borrowing tool. Friends in the finance industry may know about it. Large companies may need to borrow money on a short-term basis to supplement their cash flow when paying salaries, and they use commercial paper. So, you can see that this is a relatively frequent scenario. At the same time, the scale is also quite large. For example, Walmart, with a large number of employees globally, issues about $60 billion in commercial paper annually.

In the past, this vertical application scenario was completely handled manually in cooperation with traditional banks. It was a large-scale and important but also very boring and repetitive process. So the company we invested in automated this process using artificial intelligence. You don't even need to call it an agent because it is a very simple application of reinforcement learning and machine learning.