Alex Wu, Founder and CEO, DeepWisdom

Few founders have been chasing the same idea as long as Alex Wu. Since founding DeepWisdom in 2019, he has pursued a single goal - getting AI to build AI - first through enterprise automation, then through MetaGPT, the open-source multi-agent framework that has gathered close to 60,000 GitHub stars, and now through Atoms, a commercial platform that aims to take an idea all the way to a production-ready, deployed application.

He is also one of the more outspoken voices in the agent space: he coined the term "vibe business" to describe what comes after "vibe coding," has pushed back on the "one-person company" hype, and talks openly about "silicon-based participants" joining the workforce. We sent him a set of questions about Atoms, the multi-agent approach, and how he actually works day to day. Our questions are in bold.

Can you give us a brief introduction to yourself and to DeepWisdom - what led you from large-scale AI deployment at Huawei and Tencent, through MetaGPT as an open source project, to building Atoms as a commercial product?

When I was in high school, I became fascinated by questions about the meaning of life, humanity's place in the universe, and how we make sense of the world. I studied philosophy extensively, looking for answers, but eventually realised that philosophy was never intended to solve practical problems. So I moved on to a different question: could we build a machine capable of searching for answers on its own? Something potentially capable of solving problems that had previously been unsolvable. And that's why I chose computer science as my major.

Once I'd started university in 2010, I realised that I needed to start smaller if I wanted to make progress with the bigger vision later, so I began applying automated machine learning to quantitative trading. This was my way into using AI to build AI. Then, in 2014, I joined Tencent, where I applied machine learning to search and recommendation systems and received more than a dozen awards for my work. By 2018, as mobile internet growth slowed, I began planning a startup.

In 2019, I founded DeepWisdom, with the goal of making AI build AI. Our early enterprise solutions were difficult to scale, leading us to develop standardised AI products instead. That journey resulted in the first version of Atoms AI, and led to the creation of MetaGPT in 2023. Since then, we have been among the early pioneers of multi-agent coding, continuing to pursue the vision that first inspired me.

You've coined the term "vibe business" to describe what comes after vibe coding - AI agents acting less like tools and more like employees who execute tasks end to end. Can you walk us through a concrete example of vibe business working in practice today, ideally something one of your users has actually shipped on Atoms?

This is the full cycle of vibe business: someone gets an idea, and tells it to an agent. Then the agent, with specific skills and tools, begins to do research and build. Once the project is ready, the user can launch it with one click. Then the Marketing Module on Atoms can automatically help users improve the visibility of their project, driven by the Ads agent and the SEO agent.

So, an example would be one of our users in Kenya who uses Atoms to build a platform that educates the public about the early signs of type 2 diabetes. He developed the idea from his own experience that early detection and timely intervention are critical to preventing and managing the condition. It lowers the barrier to building an idea, testing it, and iterating on it.

Our engineer agent, Alex, helps users without a tech background to code, iterate quickly, run experiments, and build MVPs. But you still need humans to come up with good ideas and intentions.

Atoms is built on open models such as DeepSeek and Qwen, rather than on closed frontier models like Claude or GPT. Walk us through that architectural decision. Are you treating DeepSeek and Qwen as swappable infrastructure, or are you building deeper dependencies into specific model behaviours?

Actually, we use frontier models too. Our users can choose models themselves, either DeepSeek or Claude. It depends on their needs, so both options are there. The decision was really to provide our users with as much choice and flexibility as possible.

But yes, we are treating DeepSeek and Qwen as swappable infrastructure. Like us, they are evolving quickly, and in some scenarios they can already perform as well as the more advanced models, at a lower price.

MetaGPT pioneered an SOP-based multi-agent collaboration model - agents with defined roles working through structured processes. The broader industry has been drifting in the opposite direction, toward single, highly capable agents with better tools (MCP, computer use, and so on). Where does the multi-agent SOP approach genuinely outperform a single capable agent with the right tools?

In business, problems are never simple or straightforward, so solutions can't be either. In most cases, you need the coordination and cooperation of people with different talents, capacities and perspectives to find answers, and you can't always easily manage that coordination when you're working with multiple single agents.

Multi-agent architecture has the ability to simulate the real way that businesses work on a problem: discussion, specifying a task, and asking people with professional knowledge to work on it. For me, that seems like the more obvious solution. It's cohesive and comprehensive, but it's also effective.

You've publicly pushed back on the "one-person company" narrative - arguing that AI doesn't reduce competition, it intensifies it, and that DeepWisdom itself is hiring aggressively. Yet Atoms is, on the face of it, a product that lets a solo founder build and ship a fully functional business. How do you reconcile the two?

We're building the basis for the one-person company, but this isn't new. There are millions of "one-man band" organisations already out there, and the job is not that easy to do. So we're helping them. In many cases, that will enable these solopreneurs to scale, take on employees, and do more. But to get to that stage, most businesses need people with different talents to push the boundary, and the majority of startups simply don't have the capital to do that. So we're here to help them bring something new to the world.

In my view, the future of society will include a growing number of what we might call "silicon-based participants," and companies will be no exception. Organisations will no longer consist of humans alone. They will increasingly be shaped by collaboration between humans and AI. In that context, the central question is how to meaningfully expand organisational capability, and how to enable these new non-human contributors to play a real role rather than remain at the level of supporting tools. That's what we're doing with Atoms.

We're always interested in how the people building these tools actually work themselves. What does your own AI productivity day look like? Beyond Atoms and your own stack, which AI services have genuinely earned a place in your workflow versus the ones you've tried and abandoned?

I use ChatGPT a lot because its writing is very structured and logical, and Gemini is very good at phrasing. Grok is also interesting. There is a kind of beauty in its lack of alignment. When it crosses boundaries, it can become very ineffective, but within those boundaries it is highly expressive. At times, it feels as if it knows that it is wrong, yet still "can't help itself." That experience makes me think that alignment is often necessary after all.

Looking 12 to 18 months out, what's the development at DeepWisdom, or in the multi-agent space more broadly, that practitioners should be paying closest attention to - and what's the thing that's currently being hyped that you think won't pan out?

In 12 to 18 months, our users will get the first layer of the abstraction of Atoms, which means they will get an abstraction of how to build a multi-agent system themselves.

What I worry about most is whether an organisation is truly AI-native. And that takes us back to those "silicon-based participants" and how they can be used to meaningfully expand organisational capability. We don't need to keep focusing on the all-or-nothing approach to AI; the collaborative approach is much more effective.

As for hype, there's the ongoing narrative that AI will immediately replace humans. I just do not see it that way. AI is not a substitute for people. It is changing the way intelligence and creativity are created and judged, and how people work. Over time, we will adapt to it and reshape the workflow itself. So for me, the real question is whether we can build effective human-AI collaboration, closed-loop decision systems, and new organisational models that actually work - not what we'll all do when AI makes us redundant.

Thanks to Alex Wu for taking the time. You can connect with Alex on LinkedIn and find out more about Atoms at atoms.dev.

This article was produced with editorial assistance from EDDIE and MARVIN, AI tools used by the Conversational AI News team.