When AI Started Opening Stores

The paradigm shift from vibe coding to vibe business, and what it means for the future of work

In April of 2026, three Chinese software engineers walked into the Yiwu International Trade Market — the largest wholesale market on earth, a labyrinth of 75,000 booths selling everything from Christmas decorations to Bluetooth earbuds. They were not there to buy. They were there to recruit.

Their startup, Codeflying, had built something unusual: an AI platform that could generate an entire online store in 30 seconds. But the stores were empty, and the team had realized that the real problem was not code generation. It was supply. The merchants in Yiwu had oceans of inventory and no digital sales channels. Millions of regular people wanted to sell online but had no access to wholesale supply chains. Codeflying sat exactly between them.

A month later, “Yiwu AI Store” launched. A person with no technical background could browse real Yiwu products, describe what they wanted to sell in plain Chinese, and walk away with a fully operational business — storefront, inventory, AI sales staff, order management, the whole thing. No code. No warehouse. No customer service training. Just pick products, share the store, earn the margin.

This is not a better version of Shopify. It is a different category entirely.

The arc from code to commerce

To understand why this matters, rewind to February 2025. Andrej Karpathy, the former Tesla AI director and OpenAI co-founder, posted a term on X that instantly went viral: “vibe coding.” The idea was simple — stop writing code line by line and instead describe what you want in natural language, letting AI generate the entire application. Cursor, Bolt.new, Lovable, and Replit Agent exploded. By early 2026, 92 percent of developers were using AI coding tools in their daily work.

But a ceiling appeared almost as fast as the tools. The code AI generated worked — barely. It was functional but ugly, fragile, and full of unmaintainable shortcuts. “Sometimes when I look at AI-generated code, I have a heart attack,” Karpathy admitted. The output was demos and landing pages, not production systems you could trust with real customers.

A year later, in February 2026, Karpathy officially evolved his framework. He introduced “Agentic Engineering” — not just generating code, but orchestrating teams of AI agents that could design, review, and supervise each other’s output. The human’s role shifted from writing code to writing specifications and auditing results. It was a major step forward. But it still left a question unanswered: who runs the business after the code is shipped?

That question is where Codeflying entered the frame.

What happens when AI runs the store

Here is what the Yiwu AI Store experience actually looks like. A user opens the Codeflying app. They browse a catalog of real products sourced from Yiwu merchants — actual inventory, not stock photos with drop-shipping promises. They say something like “make me a store selling cute stationery for students.” The platform’s multi-agent system activates: a requirements agent structures the business brief, a page generation agent builds the storefront, a product listing agent pulls inventory data and creates listings. In 30 seconds, the store is live.

Then the operational layer kicks in. An AI sales assistant begins handling customer inquiries, recommending products based on shopper preferences, guiding toward purchase. An operations assistant auto-generates marketing posters and social media copy. An order management agent tracks fulfillment from the Yiwu merchant directly to the customer. The store owner never touches inventory — the wholesaler ships directly.

This is the difference between generating a website and generating a business. Vibe coding tools produce pages. Codeflying produces transactions.

The three forces driving the shift

Three things had to converge for this to work, and they happened to converge first in China.

The first is AI agent role specialization. We have moved past the era of the generic chatbot. Codeflying does not have one AI assistant — it has a sales assistant, a customer service agent, a marketing helper, and an operations manager. Each has distinct responsibilities, knowledge boundaries, and handoff protocols. This is not a clever system prompt. It is organizational design applied to software.

The second force is the explosive growth of one-person companies. China now has over 16 million solo-entrepreneur businesses, accounting for 27.4 percent of all enterprises. More than a third of new companies are founded by single individuals — up 53 percent from 2019. The core bottleneck for these entrepreneurs has never been coding ability. It is everything else: sourcing products, setting up payment systems, managing customer relationships, handling logistics. AI agents that absorb those operational burdens unlock an entirely new population of potential business owners.

The third force is China’s unique industrial geography. Yiwu, Huaqiangbei in Shenzhen, Shisanhang in Guangzhou — these are the densest supply chain clusters in the world. Codeflying’s most defensible asset is not its agent architecture, which competitors can replicate. It is the merchant relationships built through months of boots-on-the-ground work in Yiwu’s wholesale corridors. Silicon Valley’s path has been AI plus code. China’s path is becoming AI plus supply chain plus operations.

The numbers that make this real

When Tencent Cloud published its analysis of the AI agent economy earlier this year, one figure stood out: AI agents reduce the annual operating cost of a solo business from 225,000to225,000to18,000 — a 92 percent drop. At that price point, entrepreneurship stops being a career gamble and becomes an accessible side project. A student, a stay-at-home parent, or a retiree can run a small online business between other commitments, not instead of them.

Codeflying has raised a multi-million-dollar Pre-A round led by Fosun RZ Capital with participation from MiraclePlus (formerly Y Combinator China). Its founding team comes from Tencent — veteran engineers who know both software and the Chinese internet ecosystem. The bet their investors are making is not that Codeflying will build a better code generator. It is that the intersection of AI agents and Chinese supply chains creates a market that looks nothing like the one Cursor and Bolt are fighting over.

What we do not know yet

None of this is guaranteed. Codeflying faces existential platform risk — if Pinduoduo, ByteDance, or Meituan decide to embed similar AI commerce capabilities, they bring user bases and merchant networks that a startup cannot match. Supply chain depth is another open question: Yiwu is one market, and expanding to electronics, apparel, or cross-border categories is operationally enormous. Two-sided marketplaces are infamously difficult to bootstrap — you need enough users to attract merchants and enough merchants to attract users, simultaneously.

There is also a subtler risk: if AI generates similar-looking stores for similar products, individual sellers lose their edge. Differentiation might require the very skills the platform promises to make unnecessary.

But the direction of travel is clear. When Karpathy spoke at Sequoia’s AI Ascent 2026 conference, he described the next frontier as AI systems capable of autonomous action in complex environments. A store that runs itself — with AI agents handling sales, support, and operations — is precisely that kind of environment.

The infrastructure moment

There is a pattern that repeats across technology history. Building a website used to require knowing HTML, CSS, and some backend language. Now Squarespace and Wix make it a weekend task. Sending money across borders used to require a bank, a wire transfer form, and several days. Now you tap a button.

When something moves from skill to infrastructure, the market it creates is orders of magnitude larger than the market it replaced. If opening a store follows the same trajectory — and Codeflying is the first real signal that it might — then the opportunity is not incremental. It is generational.

The question is not whether AI can write better code than a human. The question is whether AI can help people who never wrote a line of code, never managed inventory, and never handled a customer complaint become genuine business operators. If the answer is yes, we are watching the opening scene of something much bigger than a productivity tool.