Allen Zhu invested in a Vibe Workflow company

"Building 'Coze for Beginners'" By Muxin Xu

When Andrej Karpathy tossed out the term "Vibe Coding" on social media earlier this year, he probably didn't expect it to become a darling of the venture capital market so quickly. Collins English Dictionary even named it the 2025 Word of the Year.

"Vibe Coding" essentially describes a transfer of power in human-AI collaboration: humans are responsible only for providing the "Vibe" — intent, intuition, and direction — while leaving the tedious process of code implementation entirely to AI.

But in the Workflow domain, this fluid collaborative experience has long been absent. Current mainstream tools (such as n8n, Dify, etc.) are still riddled with complex node configuration, parameter debugging, and API integration, which keeps many ordinary users locked out.

Refly.ai spotted this opportunity and defined it as "Vibe Workflow." Consistent with the philosophy of Vibe Coding, Refly.ai is committed to achieving "one-sentence Workflow generation" — users only need to describe their needs in natural language, and AI automatically completes the orchestration and configuration of complex nodes. This means ordinary knowledge workers who don't understand code can also enter the automation field, solving repetitive tasks in daily work through simple instructions.

Recently, Refly.ai announced the completion of a multi-million USD seed round, with a valuation approaching ten million. This round was co-invested by GSR Ventures, Hillhouse Capital's venture arm, and ClassIn. According to An Yong Waves (暗涌Waves), GSR Ventures Managing Partner Allen Zhu locked in the term sheet within a week of meeting the team.

What drove both Allen Zhu and Hillhouse to bet on this team was precisely this crew billing itself as the "strongest vibe workflow" team, and their attempt to define a "new paradigm" for AI applications — "the Canva of the AI Workflow space."

Part 01

ByteDance Veterans Turn Founders

Refly.ai founder Wei Huang is a veteran of ByteDance's workflow product line. Before founding Refly, his team had been responsible for an internal ByteDance product codenamed "Aily" (the predecessor to Feishu's intelligent assistant), and later the widely known Coze. They were among the earliest in China to explore the "LLM + Low-code + Workflow" product form.

"We started exploring AI + workflow at the end of 2022 and beginning of 2023. AI had just emerged, everyone thought low-code was going to be killed off, and our team felt the most urgency, so we invested the most," Huang recalls.

But through several years of practice at ByteDance, Huang discovered a pain point: whether traditional low-code platforms or the later AI-enhanced Coze, they remained fundamentally "tools for programmers."

"You have to write code, deal with if-else logic — this directly shuts out novice users, product managers, and operations people." Huang found that even conceptually advanced products like Feishu's intelligent assistant often ended up usable only by implementation specialists in practice.

This triggered his reflection: workflow in the AI era shouldn't be patching up old flowcharts, but rather required an entirely new, AI Native form.

At the end of August this year, Huang completed the financing close, and the team rapidly expanded from an initial 2 people to 14. In terms of composition, this is a team that combines technical capability with commercial execution — they had previously worn multiple hats in R&D, product, and sales at ByteDance. In their view, what gets enterprises to pay isn't just the tool, but "advanced processes."

The current Agent market presents two extremes.

On one end are general-purpose Agents like Manus and Genspark. They're like a magic dialog box — you input instructions, it gives you results. Huang compares them to a "black box": "Although the demos are fancy, the process is uncontrollable and the results unstable. It's like autonomous driving where you're stuck in the back seat helplessly watching."

On the other end are professional Workflow tools like n8n and Dify. They're powerful but complex, like a manual transmission race car that requires you to understand APIs and parameter configuration — extremely high barrier to entry.

What Refly wants to build is "intelligent assisted driving" — Vibe Workflow — somewhere between the two. On Refly's platform, users don't need to configure complex nodes from scratch. Every node is a packaged, powerful Agent. Users only need to describe their needs in natural language (the "Vibe"), and AI automatically generates a complete workflow. More importantly, this process is "white-box."

"We package the workflow into a simple, good-looking landing page. Users only need to fill in parameters, or not even that, to run it," Huang explains. In this process, users retain the right to intervene at any time. You can watch it run, pause when something feels off, modify, or manually take over. This addresses the biggest concern in current enterprise AI adoption — controllability.

"We're not trying to replace people, but to let people assemble AI capabilities like Lego bricks," Huang says.

Part 02

Winning the Creator Economy First

In Refly's strategic map, Huang has his own plan for "who will use this."

The first wave of users are tech early adopters "escaping complexity." They may have experience with n8n or Dify but are frustrated by the complex setup barrier. Refly's "one-click migration" feature allows them to directly import their originally complex flows and run them on a more lightweight platform.

The second wave of core growth is precisely targeted at "self-media" operators and content creators.

Huang's team keenly identified two pain points: first, models iterate too fast — creators want to use Gemini today, try Claude Opus tomorrow, and using them in isolation is inefficient, so they urgently need to chain them into workflows; second, the pressure of "chasing hot topics" — through Refly-built workflows, creators can automatically capture trending topics across the entire web and batch-generate articles or podcast content in their own style.

"If a self-media creator finds this tool valuable and starts using it, it's essentially covering and radiating out to their fan base as well." This high-ROI amplification effect is key to Refly's early user acquisition.

Huang gives the example that AI domain KOLs like "Digital Life Kazik" could completely turn their "article-writing workflow" into a template on Refly. Fans wouldn't just read the articles — they could pay to run this template and generate an article in that creator's style. "It's a bit like a text-based filter. Fans run your template, get results, and you make money too."

However, what users purchase isn't just the template, but a possibility — whether they too can become "Kazik" after this payment? Just as this ByteDance veteran Huang observed, many enterprises bought Feishu not just because it was a decent tool, but because of a hypothesis: "Will buying Feishu make us an advanced management enterprise like ByteDance?" But whether creators like Kazik have the motivation to open-source their templates remains to be seen.

Currently, Refly has already proven out several typical scenarios, such as financial investment research data monitoring, self-media multi-source information aggregation and topic generation, etc. Huang told An Yong Waves: "No matter how much tech concepts get hyped, it ultimately has to reach the masses. If you're only serving hundreds of thousands of professional users, it's never a mass-market product."

Image source | Unsplash

Recommended Reading

The flow of money, the rise and fall of people