Beyond Vibe Coding: Building Your Entire Business with AI

Ben Lorica

Ethan Ouyang on Vibe business, multi-agent orchestration, and shipping revenue-ready products.

Subscribe: Apple • Spotify • Overcast • Pocket Casts • AntennaPod • Podcast Addict • Amazon •  RSS.

Ethan Ouyang, Technical Lead at Atoms, joins the podcast to discuss the rise of “vibe business”: multi-agent AI systems that can research a market, design a product, build and launch it, and iterate toward revenue. The conversation why many Chinese AI startups are launching in the U.S., what U.S. founders demand from AI products, and why reliability requires orchestration, evaluation loops, and human-in-the-loop approvals. Ethan explains how modular backends (auth, payments, databases, deployments, SEO), side-by-side model “race mode,” and what it means to hand off AI-generated products to a human engineering team.

Subscribe to the Gradient Flow Newsletter

Interview highlights – key sections from the video version:

Jump to transcript

Related content:

Support our work by subscribing to our newsletter📩

Transcript

Below is a polished and edited transcript.

Ben Lorica: Today we have Ethan Ouyang, Technical Lead at Atoms, which you can find at atoms.dev. The taglines on the website are “Turn ideas into products that sell” and “AI employees that validate ideas, build products, and acquire customers in minutes without coding.” With that, Ethan, welcome to the podcast.

Ethan Ouyang: Thank you, Ben. Thank you for having me.

Ben Lorica: Before we dive into Atoms, I have a few high-level questions. We are seeing a wave of Chinese AI startups, like Manus and your team, launching in the US. What is the reasoning behind this? Is the Chinese market so competitive that startups want to explore the West instead? Why are Chinese startups suddenly launching in the US in particular?

Ethan Ouyang: I think the US is the natural place to launch AI products at this time for a couple of reasons. First, the US is where people actively build and monetize on the internet. We have creators, indie hackers, solo founders, and small businesses who are constantly experimenting by launching websites, products, stores, and communities. That behavior already exists, and AI simply accelerates it.

Another reason is the willingness to pay. In the US, consumers are used to paying for software that helps them create, earn, or save time. This makes it much easier to validate whether a product actually delivers value. The US market gives us very clear signals. There is also a strong flywheel here: capital, cloud infrastructure, and AI talent are all concentrated in the same place. This makes it easier to iterate quickly, ship faster, and learn from real users. It is very natural for any AI company to launch in the US to test, experiment, and seek growth.

Ben Lorica: In the past, I chaired AI conferences in China before COVID. My impression was that the market there is incredibly competitive—you have an idea, and two weeks later, you have 20 other startups to compete with. Is it less competitive here?

To explore that further, many observers feel that in China, the focus is on the application of AI—helping companies apply the technology. In the US, tech companies seem more focused on foundation models and AGI. Is the move to the US driven by lower competition, or is it because of this difference in focus?

Ethan Ouyang: Competition is fierce globally for AI, both for foundation models and at the application level. However, the US is uniquely important because it is where global standards for AI products are set—commercially, technically, and culturally. It is always valuable to prove your concept in this market. As you mentioned, the access to capital, cloud infrastructure, talent, and the atmosphere in the Bay Area—with all the hackathons and events—is very attractive. We really enjoy that environment.

Ben Lorica: Practically speaking, what do US developers and founders demand that differs from the Chinese market? Are we talking about security or privacy? What is different when you launch here?

Ethan Ouyang: I don’t see major differences in those areas. We are open to the public, we respect user privacy, and we follow all regulations. Our focus right now is simply on adopting more customers, enhancing our products, and improving the user experience.

Ben Lorica: For our listeners, the team at Atoms uses the phrase “vibe business.” Everyone has heard of “vibe coding,” but you are taking it to the next level: launching an entire business using prompts. Nowadays, the challenge isn’t just adoption; it’s reliability. How do you bridge the gap from a demo to a system that is reliable and delivers what users expect 95% of the time?

Ethan Ouyang: That’s a good point. AI is shifting from demos and concepts to execution and real outcomes. The US is the most outcome-driven market in the world. Customers don’t buy AI because it’s impressive; they buy it because it moves business metrics.

Atoms is designed to be a platform for building revenue-ready products with autonomous AI teams. Instead of just helping people write code faster, we help them make decisions, execute, and monetize end-to-end. In a single prompt, Atoms can research a market, design a product, build a system, launch it, and optimize revenue. This isn’t just a prompt or a model; it’s a complex system. It involves system design, specific algorithms, and orchestrating agents to coordinate and deliver accurately. We have spent years researching and open-sourcing our foundations, and everything was built on top of that work.

Ben Lorica: With vibe coding, you can validate the result because it’s a piece of code you can run and debug. You are talking about building the business itself, which includes sales, marketing, and product management. Which aspects are currently the most mature within the Atoms platform?

Ethan Ouyang: Our product is well-tested and verified. We use a multi-agent system where different agents are responsible for different aspects. We have product designers, researchers for our “deep research” feature, and engineers who execute the plan.

Ben Lorica: For coding, I prompt, get code, and see if it works. For product design, you usually need user testing. What is the validation or evaluation step for the different parts of Atoms?

Ethan Ouyang: Each part has an internal mechanism to run loops. We even have a “user agent” to verify the results.

Ben Lorica: But a user agent is just a virtual user. It might not be statistically representative of an actual human.

Ethan Ouyang: True. To make it 100% reliable, we have a “dogfooding” phase and a “human-in-the-loop” phase. Our engineers and designers test key features and vital behaviors to ensure the prompts perform correctly.

Ben Lorica: Can you walk us through a success story where a user used Atoms from scratch to launch a business?

Ethan Ouyang: Take a direct-to-consumer (DTC) brand as an example. A user might start with nothing more than a rough idea or a few sketches. Our multi-agent system spins up, analyzes market demand with deep research, defines the target audience, and sets pricing. It then designs the pages, builds the storefront, integrates core features like payments and fulfillment, launches SEO pages, and deploys everything. The user agent then tests the features to ensure they are properly implemented. The first result might not be perfect, but users can keep adding prompts to communicate with the agents and customize features.

Ben Lorica: If the DTC brand involves a physical item like a t-shirt, do you hook into the manufacturing supply chain?

Ethan Ouyang: Yes, that includes external tool calls. We can call APIs for companies like Printful to implement products, create SKUs, and handle previews.

Ben Lorica: You mentioned the deep research agent, Iris. What are the other key agents?

Ethan Ouyang: There is the Tech Lead, who organizes the priority and sequence of execution. Then the Product Designer designs how the website looks and how users interact with it, including the frontend, backend, and database tables. The engineers then take over the execution, running their own loops to iterate on the product. Finally, the user agents perform testing to ensure there are no fundamental errors before delivering a preview page to the user.

Ben Lorica: In the Atoms interface, can I track what’s happening? For example, can I approve the product designer’s work before moving to the next step?

Ethan Ouyang: Yes, that is a very important part of the user experience. Users want to know if things are running smoothly or if they need to provide more input. For vital decisions, we require human approval. The Tech Lead or Designer will provide a plan, and the user can edit it—for example, by asking to include a specific payment feature—or hit approve to let the engineers start.

Ben Lorica: You recently introduced a backend. Is that a separate offering, and what exactly is it?

Ethan Ouyang: We have been iterating on the Atoms backend for a couple of months. It is now in a modular format, making features flexible so they can be plugged in or unplugged depending on the user’s application. This includes databases, authentication, payments, deployment, and SEO pages. We are also adding features like shopping carts for e-commerce.

Ben Lorica: These features seem geared toward developers. A non-technical person might not know what they need from a backend. Does this give technical users more fine-grained control?

Ethan Ouyang: Exactly. We are open to all customers. Many users don’t have a technical background, and that’s fine because our agents handle the backend. They just need to review the plan and click approve.

Ben Lorica: Is Atoms currently geared toward online businesses and e-commerce? You mentioned Stripe integration and SEO. Do you also integrate with ad platforms?

Ethan Ouyang: We have our own infrastructure for databases, payment features, and a unified OAuth system. In the future, we may introduce recommendation features as well. We keep iterating based on priority. Our core users are solo founders or operators in small-to-mid-sized businesses who have ideas but are constrained by execution.

Ben Lorica: You also launched “Race Mode.” Is that about optimizing latency?

Ethan Ouyang: Race Mode allows you to run multiple agents or products simultaneously, each powered by a different foundation model. This lets the user compare them. Even if models have similar performance, they might have different “tastes” or styles. It gives users a chance to see which style they prefer for their product.

Ben Lorica: So if I’m in the design phase, I can use different models side-by-side to evaluate them?

Ethan Ouyang: Exactly.

Ben Lorica: Do you mainly target open-weight models, or can users pick specific ones like Gemini, Claude, or OpenAI?

Ethan Ouyang: We have most of the popular standard models in our system. Most of the time, users can use “auto mode,” but they can explicitly pick a model if they want. Claude is very popular for coding agents, while Gemini or OpenAI are great for multimodal capabilities.

Ben Lorica: What is the most technically complex agent you provide? Is it the deep research agent?

Ethan Ouyang: Autonomy isn’t about a single model; it’s a system problem. The most difficult part is coordination—context engineering and system design to ensure low latency and high accuracy. We have published papers on this execution layer.

Ben Lorica: One challenge with multi-agent systems is inter-agent communication. One agent might lie to another or fail to complete a task. How do you solve that?

Ethan Ouyang: That is where the algorithms come in. We use theories from our published papers. Most of the time, the agents run in loops using algorithms like ReAct, where they act, observe, and think. It’s not about whether an agent “lied,” but whether the result can be validated. Once the outcome is confirmed to be correct, we move to the next stage.

Ben Lorica: It sounds like a version of Shopify or WordPress, but entirely built with AI. Building and launching is one thing, but businesses need to grow, be maintained, and be secured. How do you handle long-term iteration?

Ethan Ouyang: The real shift here is speed and optionality. When execution becomes near-instant, markets can be tested in days instead of quarters. That fundamentally changes who gets to build companies. It makes it easier to get started, validate an idea, and refine it if it’s promising. If it doesn’t work, the risk and cost were low.

Ben Lorica: How customizable do you plan to be? Can I choose my own payment system or database?

Ethan Ouyang: There are tradeoffs. We are building a general, integrated platform. Supporting every possible database adds layers of abstraction that can increase difficulty and decrease stability. We evaluate what users really want every day. For now, we provide standard end-to-end solutions so users don’t have to worry about technical choices. That is the sweet spot for our users.

Ben Lorica: If I’m using your deep research agent, Iris, but I also have results from Gemini or OpenAI’s research tools, can I blend them?

Ethan Ouyang: Yes. You can add that data into the prompt to provide specific directions or more data to our deep research system.

Ben Lorica: Regarding long-term ownership: if an AI-generated app becomes successful and I hire engineers, can I export the code to GitHub to maintain it manually?

Ethan Ouyang: Yes, that is doable.

Ben Lorica: What is on the near-term roadmap for Atoms?

Ethan Ouyang: We are focused on growing the team and iterating on the product. We want to understand exactly where users feel stuck so we can polish those parts of the experience. We are also planning to introduce that recommendation feature for the backend.

Ben Lorica: Who is the dominant user profile right now? Is it a product manager or someone from a sales/marketing background?

Ethan Ouyang: It’s solo founders. Many come from sales or marketing. They have the resources and know what features people want, but they don’t have the technical background or the budget to hire a full team.

Ben Lorica: If I come from sales and don’t know anything about product design or PRDs (Product Requirement Documents), how do you fill those knowledge gaps?

Ethan Ouyang: The key is making the product user-friendly for people without that knowledge. We explain things in simple terms and communicate progress clearly. Users don’t need to write a “perfect” prompt. Our system handles the difficult parts through context engineering and system design. “Less is more” for the user because our agents take care of the complex requirements.

Ben Lorica: You mentioned that much of this is based on open research. How much of it is open source?

Ethan Ouyang: We have an open-source project called MetaGPT. People can Google that to see the core of how we originally designed our multi-agent systems.

Ben Lorica: And with that, thank you, Ethan.