Original podcast link: From Undervalued to Star Project: What Did Dify Get Right? | Interview with Dify Founder Luyu
(Note: The following content is an AI-generated summary based on the original transcript. Minor discrepancies may exist. Please feel free to point out any inaccuracies.)
Qukai: Looking back over two years—from 2023 to 2025—what were Dify’s key milestones in market and technological evolution?
Luyu: In 2023, we started with “visual Prompt Engineering” and among the world’s earliest visual “RAG” tools, paired with a user-friendly backend service architecture and interface, enabling rapid product understanding and adoption.
In the second half of 2024, serious enterprise (2B) users began entering the market in large numbers, with clear willingness to pay and production-grade requirements. Dify launched “Workflow,” built a plugin ecosystem around complex applications and intelligent processes, and focused on solving model “hallucinations” and robust integration with enterprise tools, data, and APIs. By 2025, multimodal and open-source models had significantly advanced. Early assumptions such as “model neutrality” and “necessity of middleware” were largely validated. Market focus shifted from “Prompt” and “RAG” toward broader “context engineering.” Key technical variables stabilized, and the overall architecture entered a trustworthy phase suitable for 3–5 year investments.
Qukai: What is the fundamental difference between Dify and LangChain? Why can both coexist long-term?
Luyu: The core difference lies in user positioning. LangChain targets engineers with strong coding skills, helping them accelerate development and debugging via tools like LangSmith. Dify, by contrast, caters to non-technical or even low-technical users—office workers—who can directly build SOPs and intelligent workflows using “Workflow” and “Agent” features.
If we place users along an “engineer spectrum,” LangChain sits far left as a hardcore engineering tool; Dify is in the middle but continuously moving rightward, aiming to empower non-engineers to produce production-grade applications. OpenAI’s AgentKit this year targets the middle segment—still technically capable but less hardcore.
Qukai: How do Coze, GPTs, and Dify relate and differ?
Luyu: From the start, we never considered them the same category. GPTs and Coze aim to let the general public create simple bots and enter a UGC-like marketplace. But most people lack the four scarce capabilities needed to build “production-valuable applications”:
- Hard tech skills (algorithms, cost, hardware)
- Exceptional interaction and creative design
- Unique data moats
- Deep understanding of business processes (SOP)
Among these, only “process capability” is broadly transferable; the other three are hard to replicate.
Dify places LLMs into the enterprise “fish tank,” connecting talent, tools, data, and processes to build “deployable systems,” not a UGC bot marketplace. GPTs have played a vital role in market education but have yet to form a true production-value ecosystem.
Qukai: How do you view n8n’s rise? Is it complementary or competitive to Dify?
Luyu: n8n began as an open-source alternative to Zapier, emphasizing low cost and data sovereignty. Over the past few years, it built a massive ecosystem of connectors and templates, amplified by strong marketing, elevating its brand and popularity.
In user perception, n8n and Dify do overlap: n8n excels in general automation and third-party connectors, while Dify is stronger in LLM-native problems like RAG, multimodal context handling, and scalable AI components. Dify is an AI-native engineering system that delivers “Agents, Bots, and Applications” end-to-end—from design to deployment—while n8n struggles to deliver full production-grade solutions alone. In real enterprises, the common approach is a “combo”: use n8n for general automation, Dify for complex semantics and reasoning.
Qukai: Will OpenAI’s AgentKit kill low-code and Workflow tools?
Luyu: This “either disrupt everything or be worthless” binary view usually comes from people who haven’t used these products at scale. AgentKit is still very early; in the next 6 months, it’s unlikely to pose structural threats to mature 2B products. Even if everyone has “drag-and-drop workflows,” underlying product positioning, ecosystem building, and engineering cores differ vastly. 2B scenarios demand long-term trust, open ecosystems, model neutrality, and controllable compliance. Tying your tech strategy entirely to one model and its suite carries high strategic risk. Most enterprises will choose “multi-standard middleware sockets” rather than betting everything on a closed stack.
Qukai: Why doesn’t the critique “middleware is too thin” hold water? Where exactly is engineering expensive?
Luyu: Engineering itself is what’s truly expensive. To serve real users and specific business contexts, you must make correct layering and abstraction decisions—distinguishing what’s mutable vs. immutable, finding the right granularity balance: too fine-grained reverts to reprogramming; too coarse can’t handle business complexity. These judgments come from long-term real-world磨合 (refinement) and technical tuition. We observe that 80% of AgentKit’s code may be generated by models like Codex, but “auto-generation” can’t replace engineering systems honed over years in real customer scenarios. Dify’s components are extensible and hot-swappable, maintaining version compatibility during upgrades—developers can adopt new tech without sacrificing existing assets or stability.
Qukai: With AI coding and automation advancing, is Workflow still necessary?
Luyu: Not only is it necessary, it will persist long-term. Enterprises demand extremely high predictability in production processes and reasoning logic—often requiring >95% reliability. Complex tasks must be broken into multiple “checkpoint nodes” for human-AI collaboration and quality control. “Single-node black boxes” are rarely acceptable in production; users need to insert human feedback and audits at critical points. Dify’s path starts from “high-code, high-reliability,” gradually evolving toward smarter, more automated systems. Along the way, we deliver “production-ready” tech—users experience it like a car that never breaks down, not a magical Bot that surprises but disappoints upon real-world deployment.
Qukai: Have you ever misjudged the pace of Agent evolution?
Luyu: Model capability leaps are inherently unpredictable. Several key advances did accelerate Agent development beyond early expectations. Some companies chose to heavily bet early on immature tech—a “waiting for the rabbit” strategy—which explodes once models mature. This approach can yield viral products in the short term.
We avoid such extreme risk preferences, not betting heavily on single inflection points. But we acknowledge this strategy can be effective during certain windows.
Qukai: If you could only change one thing over the past two years, what would it be?
Luyu: Team building should have been more aggressive. From early to late 2024, business grew exponentially, but headcount lagged, making delivery extremely strained. We assumed and invested in global and multi-office collaboration infrastructure and culture from day one, but we could have been more decisive in scaling headcount and preparation.
Qukai: Regarding open-source and globalization, why did Japan become a viral market? What did you specifically do?
Luyu: The key was “open-sourcing early enough.” Japan has long faced tech talent shortages while valuing processes highly. Dify felt to them like Excel did to accountants decades ago. In 2023, we provided Japanese support early; the community spontaneously exploded in April–May. During that time, Dify appeared everywhere—in TV shows, bookstores, cafes. Initially, we had zero employees in Japan; all traffic and enterprise inquiries came from organic community growth, after which we built a ~10-person local team.
Open-source here solved trust issues and accelerated diffusion and “de facto standard” formation—our most fundamental global strategy decision.
Qukai: What gives you confidence to plan a 3-year tech roadmap?
Luyu: Today’s key technical variables are largely stable; models are no longer completely opaque black boxes. We don’t predict specific model capabilities but focus on “how to help everyone use electricity well,” not “how to build power plants ourselves.”
But for fine-grained real-world components like Memory, MCP, etc., we believe ecosystems must co-build. Dify’s 1.x release, in a sense, declares two things:
- The external tech environment has entered a relatively stable phase.
- Our architecture is trustworthy over a 3–5 year horizon.
Qukai: From your AI worldview, why do you insist on “neuro-symbolic AI”?
Luyu: LLMs belong to the “neural network” paradigm, excelling at association and expansion. But the human brain is efficient because it intrinsically possesses “symbolic ability” and causal reasoning, enabling rapid decision-making like a dynamic decision tree.
We firmly believe that truly scalable, production-grade Agents require a hybrid “neural + symbolic” system: leveraging neural networks’ associative richness while using interpretable logic and constraints to ensure reliability and controllability. This isn’t an engineering fantasy—it’s grounded in long-term bionics and cognitive science research. The narrative that a single Transformer can “solve everything” is implausible under real-world constraints.
Qukai: In a multi-model environment, where does model neutrality’s long-term value lie?
Luyu: The market has fully validated this. Enterprises generally dare not tie their tech strategy entirely to one model and its suite. Neutral middleware acts like a “universal socket,” balancing compliance and openness, making long-term tech investments safer and avoiding single-point strategic risks. In this landscape, our competitive edge against model vendors becomes clearer.
Qukai: Can you give a concrete engineering example—how do you ensure compatibility during version upgrades and component extensions?
Luyu: In Dify, key components like Workflow and RAG are modular. Users can use our native kits or replace them with self-built components, supporting hot-swapping and independent upgrades. During the 1.0 to 1.9 upgrade, user-built components remained functional. We design to balance “introducing new capabilities” and “preserving old assets,” ensuring upgrades don’t become high-risk “business restarts.”
Qukai: Community Edition upgrades are “fast but buggy,” while Enterprise Edition prioritizes stability—how do you think about this trade-off?
Luyu: In open-source collaboration, with the core team plus over 1,000 contributors iterating, achieving deep and broad automated test coverage is an extremely difficult engineering challenge. Stability and technical aggressiveness must be balanced. Enterprise customers get more stable update rhythms and stricter testing; Community Edition moves faster on new features—e.g., MCP support lands first in Community.
Qukai: Will the future relationship between SaaS and enterprise self-built systems evolve into “fast-fashion SaaS” being replaced by internal AI?
Luyu: At the human-machine interaction level, replacement is possible; but at the software engineering level, many “stable data structures” won’t change easily. Take financial systems: invoice fields, sensitivity levels, long-term retention rules—these can’t be arbitrarily moved. We believe future efforts will focus on defining “immutable structures and outcome standards,” while AI handles dynamic customization of processes and interactions above these standards. Thus, “fast-fashion SaaS” replacement is only half-true: the standard layer still needs professional software; the non-standard “glue layer” is driven by AI and Workflow.
Qukai: In real enterprise applications, how are standardized vs. non-standardized problems divided? Where does Dify excel?
Luyu: The most common and thorniest parts in enterprises are often non-standardized. Vertical SaaS has already optimized standardized, repeatable problems. What truly troubles enterprises is integrating existing systems, processes, data, and tools into complex collaboration networks. Previously, this required expensive system integration and hand-coded engineering; now, LLMs and AI collaboration canvases dramatically compress this cost. Dify’s value lies in handling this non-standard “glue layer.”
Qukai: Will Dify’s final form be an “operating system”?
Luyu: Rather than an OS, it’s a new organizational collaboration paradigm.
We aim to create an intelligent platform as the enterprise’s collaboration hub: plug in “atomic capabilities” as modules, incorporate humans as feedback nodes, unify process and dashboard views. Ideally, at 9 a.m., a person opens their phone to see pending reviews and feedback tasks, then enters a process to instantly view the entire production panorama. This form is already very close to an “enterprise OS.”
Qukai: As AI capabilities overflow, how will human-model collaboration evolve?
Luyu: Today’s open-source models are already “luxuriously sufficient” for most people. The real challenge lies in:
- Clearly articulating problems and context.
- Learning to extract answers using the right “evidence passwords” in vast data spaces.
Many lack this “path-finding ability” and questioning framework. We aim to provide “treasure map”-style interaction maps via Dify, guiding users steadily toward correct answers. In human-AI collaboration, humans act as “real-world sensors,” continuously feeding field data and feedback back to models.
Qukai: When models become symmetric, what remains as an organization’s true competitive edge? How does Dify preserve these?
Luyu: When model capabilities converge, what remains are “asymmetric assets”: values, decision mechanisms, attention allocation, and unique process assets. We’re exploring how to solidify these unique contexts and process logics via Memory, Workflow canvas, etc., forming reusable “organizational memory and differentiated reasoning.” Thus, a system starting at 60 can evolve to 70, 80 through continuous feedback and context expansion; top individuals and organizations maintain irreplaceability at 90, 100 via asymmetric capabilities.
Qukai: Why did you propose the role of “Chief Context Officer” (CCO) in organizational building?
Luyu: The biggest problem with today’s Agents is disordered, noisy context organization.
We aim to turn all organizational information into an efficient information bus, setting clear access orders and governance mechanisms so Agents can instantly attach global context and precisely locate relevant nodes when answering questions.
From this perspective, every enterprise should have a “Chief Context Officer” to systematically govern this context bus—it’s one of the core future organizational roles.
Qukai: Studies suggest “95% of enterprise AI pilots fail.” What’s your take on this phenomenon, its root cause, and opportunity?
Luyu: This isn’t a model capability issue but a huge gap between “tools and organizational learning ability.”
The next key opportunity isn’t building stronger models but “building bridges”—creating infrastructure, human-AI interaction paradigms, and efficient workflows that bridge this gap, making “how to collaborate with AI and efficiently extract data to complete work” a widespread skill. The biggest opportunity in the next few years lies in solving AI’s “last-mile” problem.
Qukai: Who can truly use AI well, based on your observations? What advice do you have for enterprises?
Luyu: Currently, only a few can use AI well. They constantly operate near model boundaries, maintain strong curiosity, and frequently experiment and iterate. We even track developers’ model consumption—distribution is far more extreme than the 80/20 rule. Our advice: enterprises should identify and cultivate these “super individuals” early, enabling them to collaborate with ten or more AIs, turning context and processes into replicable organizational assets.
Qukai: How do company culture, benefits, and global collaboration serve these “super individuals”?
Luyu: We uphold a culture of “transparency to investors, trust first.”
For investors, we allow access to internal documents and data, avoiding over-polished reports.
For employees, we provide a $1,000 credit card upon onboarding, preferring post-hoc review over pre-emptive restrictions. We heavily adopt U.S. SaaS and global toolchains, emphasizing asynchronous collaboration and multicultural integration. As a result, with 70–80 global employees, our attrition rate remains below 5%. This environment lets top talent unleash productivity in high-trust, low-friction settings.
Qukai: In work policies, you oppose both 996 and rigid 8-hour systems?
Luyu: We oppose the mental burden of long overtime and the mechanization of creative work into rigid 8-hour quotas. Creators’ best state is flow—intense focus, losing track of time, rapidly converting inspiration into output. Rules should serve maximizing overall output, not constraining everyone with uniform time quotas. For a team, goal alignment and member happiness matter more than rigid systems.
Qukai: As a founder, how do you maintain happiness and creativity?
Luyu: Founders must actively make themselves happy, adjusting mind and body to a creative state so good ideas emerge naturally in relaxed environments, not forced under pressure. Don’t let external expectations and stress crush you. Sometimes, just walking around Stanford’s campus, inspiration arrives unexpectedly. Organizational management is similar. When you’re merely carrying burdens, you easily lose the initial impulse to “change the world.” Happiness is a prerequisite for sustainable creativity.
Qukai: In your view, what’s Dify’s long-term label? How will your strategy evolve in a maturing market?
Luyu: In mature markets, every product needs a heavy, decade-long label.
For Dify, we’ve always insisted on “engineering-first, long-termism, stability, reliability, open collaboration.” Rather than asking what we’ve changed, ask what hasn’t changed—these constants are the true cycle-proof core.
Qukai: Final question: Which roles are you currently hiring for? How can candidates apply?
Luyu: We’re hiring in China, Japan, and Silicon Valley for Product, Marketing, Backend Engineering, and CCO roles. Follow Dify’s official WeChat account for job postings, or apply directly to: joinus@dify.ai
We apologize for the inconvenience and are actively contacting support to resolve the issue 