Creating Agents with Agents? A Conversation with Yunqi Capital's Chen Yu and Creao AI's Cheng Kai on the New Agent Tech Ecosystem | Yunqi Attent!on Podcast

The rapid rise of Manus and Genspark has made Agent one of the most anticipated AI application forms for the general public in 2025. Meanwhile, the Agent development ecosystem has entered a stage of stronger compatibility and easier deployment, driven by new frameworks like MCP and Agent-to-Agent.

In this episode, we focus on the topic of Agent. We've invited Chen Yu, Partner at Yunqi Capital, and Kai Cheng, Founder and CEO of Creao AI, to discuss the evolving AI Agent technology ecosystem, the shifting center of gravity in Agent "infrastructure," Creao AI's practice of "using Agent to build Agent," and the future of human-AI symbiotic collaboration — all starting from the hot topics in AI tech in 2025.

Scan the QR code above or follow "Attent!on" on Xiaoyuzhou to listen to this episode

Guest Introductions

Chen Yu, Partner, Yunqi Capital

Deeply rooted in frontier technology sectors including AI and embodied intelligence. Representative investments include MiniMax (large model unicorn), DeepRoute.ai (leading autonomous driving company), and PingCAP (open-source database unicorn).

Kai Cheng, Founder & CEO, Creao AI

Master's in Statistics from Columbia University, Bachelor's in Mathematics from University of Toronto, studied under Turing Award winner Geoffrey Hinton. Formerly Mathematical Scientist at iPerceptions, where he developed machine learning algorithms for Dell, NASA, Comcast, and other companies during his time in North America.

Creao AI is a startup focused on AI agent infrastructure, providing enterprises with a fully AI-driven Agent operating system that helps business clients achieve automated Agent tool construction and management, facilitating efficient Agent deployment and application.

Linda, Managing Director, Yunqi Capital (Host)

Timeline

Part 1: Recent AI Hot Topics

02:22 Technical side: Functional updates across major models essentially reflect leaps in underlying model reasoning and planning capabilities, laying the foundation for significantly improved Agent generalization

05:10 Demand side: Enterprise awareness, acceptance, and understanding of AI is accelerating in tandem with improvements in large model reasoning and planning capabilities

06:59 Technical ecosystem: Infrastructure-level innovations such as Agent-to-Agent frameworks are driving enterprise deployment of Agents

Part 2: 2025, How to Build AI Agent "Infrastructure"?

09:40 A layman's definition of Agent: a "digital employee" that produces real-world impact

11:28 As Agent development difficulty continues to decrease, what value do infra tools provide?

13:19 "Stability" is a strong demand for enterprise Agent applications — even the simplest Agent needs a stable environment to run

16:41 Workflow or Agent? Chen Yu: not mutually exclusive; workflow is a more stable way to open up Agent capabilities

18:23 The underlying evolution of infra behind MCP and Agent-to-Agent: empowering large model training and upper-layer applications

Part 3: Using Agent to Build Agent: Creao AI's Practice

20:58 Creao AI's PMF exploration: helping enterprises build Agents with Agents

24:45 Creao AI client case: Agent + CRM

28:45 Changes from the demand side: more and more non-technical personnel also have development needs

30:59 The "mutual selection" between Yunqi Capital and Creao AI

Part 4: Brainstorming: The Future of Agent-Human Collaboration

34:07 AI's role: amplifier of human creativity, accelerator of technological exploration

36:18 When AI gains autonomous consciousness, humans will need to treat it as an equal species and communicate with it as such

37:30 Avoiding rigidification of AI tools: behind the "AI-ness" actually lies human behavior patterns

Below is an excerpt from this episode's podcast (some text has been edited and organized)

Recent AI Hot Topics:

Application Leaps Driven by Large Model Evolution

Linda, Host

From your gut feeling, what are one or two AI-related developments in the past month that have genuinely excited or surprised you?

Chen Yu, Partner, Yunqi Capital

Actually, the news looks pretty similar every month — GPT released some new version, Gemini released some new version. Each new release comes with enhanced large model capabilities, and with these capability improvements, new applications keep emerging.

Many applications can only be built once the large model has the corresponding functionality. For example, why didn't an Agent system like Manus come out two years ago? Because two years ago, large models' reasoning and planning capabilities were far from meeting the requirements for creating a general-purpose Agent. If you had tried to build an Agent system two years ago, implementing many things yourself, you'd find after two years that much of your work was essentially wasted. Because large models can now help you complete those tasks much better. And large models do this in a more generalized form. It's not about solving a specific problem in a targeted way, but rather, like the human brain, having a kind of generalization. No matter what problem comes in, or what environment it faces, it can handle it relatively easily.

So AI applications are far from reaching a point where they're finished or no longer advancing. Because large models themselves are still evolving, and as they evolve, AI applications have unlimited possibilities.

Linda, Host

Kai, as an entrepreneur, from the perspective of industry dynamics and client interactions, what's your take?

Kai Cheng, Founder & CEO, Creao AI

Since we serve mostly enterprise clients, I mainly look at it from the enterprise perspective. From March until now, I've observed a fairly noticeable trend: enterprises' understanding of AI is undergoing relatively rapid change and improvement. Beyond DeepSeek's widespread adoption, another reason is that in the traditional software technology era before 2023, the technology cognition cycle was relatively long — generally it would take about a year for enterprises and individuals to gradually recognize or accept technological advances. But after 2023, the pace of technological evolution has been extremely fast; everyone is experiencing what you might call accelerated cognition.

However, for enterprises, while all employees and management are cognitively catching up faster and faster, how to actually apply the latest large models or AI technology to their real business environments remains a relatively complex and somewhat uncertain matter. But starting from last year, as large models' reasoning and planning capabilities significantly strengthened, enterprise users have begun to understand how to combine AI with their own scenarios, or what kind of AI capabilities they actually need to generate value for their business. So this is actually a fairly significant change from the client side.

From a technical perspective, although as Yu just mentioned, from March until now we're all probably in the same state — every day discovering new Agent products being released, or new AI capabilities being rolled out, or technologies that completely overturn previous approaches, like image generation and so on. But what actually excites me more is frameworks like Google's recently launched Agent to Agent, which represents an innovation at the infrastructure level.

Because most people currently focus on the iteration and progress of foundation models or Agent capabilities themselves. But honestly, without better technical ecosystems, especially infrastructure-level innovations to support the deployment and application of Agents, enterprises will find it difficult to truly and efficiently utilize Agents. Frameworks like Agent to Agent show me the possibilities of how more Agents could land in enterprise settings in the future.

2025, How to Build AI Agent Infrastructure?

Linda, Host

In simple, accessible terms, what is an AI Agent in your eyes?

Chen Yu, Partner, Yunqi Capital

I think investors tend to describe things in the most abstract language possible. In the most abstract terms, it's a digital employee that can autonomously help you complete the tasks you assign according to your instructions — that's it, simple as that.

Kai Cheng, Founder & CEO, Creao AI

One is what Yu just mentioned — the concept of a digital employee. Secondly, the biggest difference between an Agent and the AI models people have encountered before is that it can have actual actions and connections with the external world. AI models themselves cannot actually connect with the world. But the difference with Agents is that after reasoning out a result, they can actually call tools in the external world and have an impact on it. This is probably why Agents can be treated as digital employees — because if it's an employee, it must be able to produce certain actions or changes in the world. This is probably an important aspect of Agents.

Linda, Host

Kai, your product is very much about Agent system design and optimization. With Agent development difficulty decreasing today, what value can you provide?

Kai Cheng, Founder & CEO, Creao AI

Indeed, as Linda just said, from today's technical environment, building an Agent system or application is much less difficult than before. Even many relatively simple workflows or Agent systems, from design to coding to deployment — the entire process can be fully completed by experienced developers in 1-2 days. But Agents as systems don't actually exist in isolation; they require the stable operation of the entire system. This depends on the construction of the entire environment where the Agent operates, including the continuous iteration and improvement of that environment itself.

Let's take a very common, hot example right now: the Web research Agent. Everyone knows what its process looks like: first crawl web pages to gather all the information, then build a website. This workflow logic is very simple today, but if we want this Agent to run properly, we also need capabilities like web crawling, PDF generation, generating visual charts or interactive web pages, and so on. Any of these capabilities, or any aspect of the environment, if it becomes unstable or has even a small problem, will cause the entire Agent system to fail. So a successful Agent system actually depends on a very stable environment. The Agent itself and its environment include task logic, callable tools, the knowledge needed when processing these business flows, and server environment setup. Each part needs to be compatible with the others — it's a very complex environment construction process.

So to put it simply, the core value we provide is using AI to help enterprises build the entire AI system and the environment it needs, and then using AI to enhance the Agent's own capabilities.

Linda, Host

In the future, will competition among Agent products shift toward automatic construction of Agent systems? There are two voices in the market now — one believes that workflow actually limits model capabilities, and true Agents don't need preset workflows; their capability is the natural inference of the model itself. Can the workflow framework be understood as meaningless?

Kai Cheng, Founder & CEO, Creao AI

Right now, when we provide products to enterprises for building Agents, if we want more stability, it might exist in the workflow form that Linda mentioned, because workflows are more stable. But if we want it to be more like a human, providing relatively creative ideas or approaches to handling tasks, then we might make it more dynamic — and that's the concept of a true Agent system. So whether workflow or Agent, both can be built quite easily now. But if you want it to run stably or successfully, you need more things together to build the environment I just mentioned. So for enterprises, the Agent itself or the workflow itself is certainly important, but not necessarily the biggest challenge. The biggest challenge is how to make it run more stably in enterprise scenarios — that's actually the more important question.

Chen Yu, Partner, Yunqi Capital

I don't think these two are mutually exclusive. Using a real-world example: in a Chinese language exam, essay writing can be a completely unrestricted topic — write a travelogue, with no other constraints. But it's also possible to write requirements very clearly, specifying what order to follow, what to write in the first paragraph, second paragraph, third paragraph — preset in the prompt, like a命题作文 (proposition composition). Workflow is more about restricting Agent behavior, and to some extent this restriction can make the Agent more stable. Otherwise, when trying to solve problems, it will think in its own very divergent way, and might get stuck in a dead end, never solving the problem. So you can understand workflow as a kind of restriction that regulates Agent behavior.

For the full episode, subscribe to the "Attent!on" channel on Xiaoyuzhou app~