What Dify Did Right to Go from Undervalued to Star Project | A Conversation with Dify Founder 路宇
Summary
- In 2 years, Dify grew from a “visual Prompt/RAG tool” into enterprise AI infrastructure with more than 110,000 GitHub stars, ranking among the global top 50, validating its day-one choices of open source, To B and globalization. 张路宇 says serious enterprise users typically enter 6 to 12 months after a new technology emerges, and Workflow, launched in March or April 2024, arrived just as demand was forming. Around 20% of Fortune 500 companies now use Dify, while the company operates in what is essentially a “zero-sales” mode. Its most important growth asset was not a one-off concept boom, but open source’s ability to solve trust, distribution, technical standards and global reach simultaneously.
- Dify’s real differentiation is not having more features than LangChain, Coze or n8n, but occupying the midpoint of the “hardcore coder—nontechnical user” spectrum and continuing to move right. LangChain serves developers with strong coding skills who need tools such as LangSmith for debugging; Coze sits further right but lacks production depth; n8n is stronger in automation, connectors and template ecosystems, while Dify leads in AI-native RAG, multimodal context and end-to-end application delivery. “The model is like a fish”: put it in an IDE and you get Cursor, put it in an Ubuntu sandbox and you get Manus, put it into an enterprise’s people, tools, data and workflows, and “it becomes Dify.”
- The moat around an enterprise platform is not “thin”; model neutrality and the engineering required for long-term compatibility are precisely the hardest parts to replicate. Enterprises will not tie years of technology strategy to a single model vendor and would rather buy “a socket that accepts different standards.” Dify also makes Workflow, RAG and knowledge-base algorithms hot-swappable and independently upgradable, allowing internally developed suites to remain usable from 1.0 through 1.9. Zhang’s core judgment is: “Everyone thinks engineering is worthless, but we believe engineering is the most valuable thing.”
- Workflow will not disappear as Agents become more capable, because enterprises want 95% or higher reliability and unattended operation, not a one-off spectacle. Complex tasks such as finding a home, communicating and signing a lease inherently contain multiple checkpoints; the number and granularity of nodes will change with model capability, but the workflow itself comes from how humans work. Using DeepSeek’s progression from large-scale coding and extremely reliable High Code toward intelligence as an example, 张路宇 stresses that Dify must always deliver production-ready technology and hold the risk line. His deeper technical belief is neuro-symbolic AI: neural networks handle association and expansion, while symbolic systems handle logic, judgment and causal constraints.
- Dify’s most valuable enterprise use cases are precisely the “nonstandard” ones, and the company is moving from a developer tool toward an organization-wide collaboration system. Existing customers have built Workflows with 400 or 500 nodes; 张路宇 says Anker may have built more than 1,000 workflows, integrated tens of thousands of atomic capabilities and, after designing SOPs, directly assessed whether each role should be staffed by “a person or an Agent.” These nonstandard processes once required expensive IT integration; Dify now acts as the “glue” connecting legacy systems, data, people and Agents, and its future form may resemble an operating system for enterprise production.
- The growth thesis was not entirely right: Agent autonomy arrived faster than expected, while Dify was not aggressive enough in staffing for 2024 growth. Companies like Cursor took positions before models were mature and waited for the capability inflection point—a high-risk but potentially effective “waiting under the tree for a rabbit” strategy. Dify will not chase short-term fads with the same risk appetite. The more concrete mistake was that the team had only a dozen or 20 people at the start of the year, while the business grew severalfold by year-end without a matching increase in headcount. 张路宇 also admits that the company initially understood only its open-source middleware mission, not its 5- to 10-year purpose.
- As general-purpose models converge in knowledge and reasoning, 曲凯 argues that people and organizations should preserve their asymmetric assets, and asks whether Memory and Workflow can carry them. 张路宇 responds that people at 60 points are likely to be replaced, while reaching 90 or even 100 points is what makes someone difficult to replace. He also believes organizations need shared Memory and an information bus, perhaps even a “chief context officer.” Models are already good enough for most people; the key is now to provide accurate context, a map to the answer, and better decisions and attention allocation.
Deep dive
1. Three Years of Market Evolution Pushed Dify from Technical Demo to Enterprise Production
Dify open-sourced its first version in the second week of May 2023. At the time, almost any product that helped ordinary people understand LLMs could become an overnight hit. Visual Prompt engineering, a rudimentary but early visual RAG, the Backend as a Service concept and a friendly interface were already enough to create novelty.
By 2024, the market had shifted from “how to use AI” to serious production, with enterprises that had real purchasing power entering mainly from midyear through the second half. 张路宇’s conclusion was that after a new technology appears, producers typically arrive at scale with a 6- to 12-month lag: “It’s not as fast as everyone imagines. You have plenty of time to work on it.”
Dify launched Workflow in March or April 2024 and then built a plugin ecosystem around it, using deterministic processes to connect LLMs, APIs and existing infrastructure while addressing hallucinations and reliability in complex applications. In 2025, open-source models, multimodality and the model ecosystem matured further, validating early assumptions around model neutrality, middleware and coexistence among multiple models.
2. User Positioning, Not a Feature List, Explains Dify’s Coexistence with LangChain
张路宇 admits that 2 years ago, when asked about the difference between Dify and LangChain, he could only say that the experience and engineering were better. His answer now is that they serve different audiences. LangChain targets strong coders who want to accelerate development and debugging, with LangSmith as a typical companion product.
Dify has continued moving toward less technical and even nontechnical users. In Japan, large numbers of ordinary white-collar workers use it to build work SOPs and internal enterprise Agents rather than treating it as a programmer’s library.
If users are placed on a 1-to-10 scale from hardcore technologists to people who know nothing about technology, 张路宇 puts today’s Dify at around 5 and wants to keep moving right. n8n sits further left and is more technical; Coze sits further right, but “its production attributes are still insufficient.” 曲凯 notes that Dify remains at the midpoint today, and 张路宇 does not present the vision as the current state.
3. The Problem with GPTs and Coze: Creators Lack Reusable Product Capabilities
张路宇 believes OpenAI had the right idea with GPTs in 2023—“let everyone create bots”—but never clarified who could build products worth redistributing. Short videos are cheap to consume, but even the smallest Agent is still a product and must solve a real problem.
He breaks AI product competitiveness into 4 categories: hard technology such as hardware, algorithms or extremely low cost; interaction and creativity in areas such as games, companionship and language learning; special moats such as proprietary data; and an enterprise’s unique understanding of SOPs and workflows.
The first 3 types of capability are inherently scarce and almost impossible for a platform to replicate broadly; only workflows are reusable. Even if ordinary users can “make something,” they generally cannot build a good product without one of these 4 sources of value.
曲凯 asks whether this means GPTs, Coze and Dify do not compete at all. 张路宇’s answer is that, from a positioning perspective, they are not even direct competitors. They are all based on LLMs, but their creators, organizational capabilities and production contexts differ. GPTs never truly took off despite ChatGPT’s enormous distribution, but they did educate the market for Dify by showing enterprises overnight that “AI can actually be used this way.”
4. The Same Model Becomes Completely Different Products in Different Contexts
张路宇 recalls that when Copilot launched in July 2023, he was anxious for less than 2 days and never opened it again, because he quickly confirmed that the two products had different audiences and organizational capabilities. From the beginning, Dify placed models inside an enterprise’s people, tools, data and workflows rather than offering an isolated bot.
The team’s internal metaphor is: “A model is like a fish. Put it in different fish tanks and it produces different results.” Put it in an IDE and you get Cursor; put it in an Ubuntu sandbox and you get Manus; put it in an enterprise environment and you get Dify. Sharing LLM technology does not mean sharing a market.
5. n8n Is a Real Competitor, but Its Strength Lies in Automation Rather Than AI-Native Capabilities
n8n is a European company founded in 2019. It initially had nothing to do with AI and aimed to replace expensive Zapier, whose data-sovereignty concerns made it unattractive, with an open-source, neutral and low-cost alternative. Years of accumulation gave it a lead in connectors, tools and templates, while UGC marketing rapidly built awareness in recent years.
张路宇 acknowledges that the two companies do have overlapping customers, with many enterprises using both. n8n is stronger at automated connectivity, while Dify got started later; but n8n’s LLM, RAG, multimodal and context capabilities remain relatively basic, functioning largely as LangChain wrapped into a workflow, whereas Dify is AI-native at the architectural level.
Dify’s other difference is end-to-end delivery: users can create an Agent, bot or other application from the interface and deliver it to the final user. n8n cannot complete that loop, so the practical answer may not be replacement but combination at different layers.
6. The Short-Term Impact of AgentKit Is Overestimated; To B Products Do Not Replace One Another Just Because They Look Similar
In response to the recurring narrative after every OpenAI product launch that “someone has been killed,” 张路宇 calls it a false binary. To B products have long-term trust, diverse connections and existing investments behind them, and one cannot infer replacement simply because two interfaces both use drag-and-drop Workflows.
His analogy is the US e-ink market: Kindle, Remarkable and Supernote look similar, but each has built a stable audience around writing, zero latency or library ecosystems. Open ecosystems, model neutrality, engineering priorities and delivery methods are the underlying core beneath the surface.
On OpenAI AgentKit, which is still in its early stages, 张路宇 gives a clear but time-bounded judgment: “It will not pose a threat to any other product within 6 months.” This does not deny that model vendors can move down the stack; it means enterprise migration and the rebuilding of trust will not happen overnight.
7. Model Neutrality Makes Dify an Enterprise “Socket” Worth Holding for the Long Term
The enterprise technology decision-makers 张路宇 speaks with treat middleware and development platforms as multiyear infrastructure, scrutinizing compliance, openness and compatibility. They are more cautious about model selection because no one wants to bind its models, development kit and full technology strategy to a single vendor.
His analogy is that if what you buy is “a socket,” you can plug in different standards in the future, making it harder for the technology investment to go wrong. 曲凯 cites shifts among model providers in US AI coding and the risks of single-vendor dependence, which also explains why enterprises need the option to use multiple models.
Dify chose open source, To B and globalization from day one, then derived an open ecosystem, model neutrality and engineering-first approach from those choices. 张路宇 believes these mutually reinforcing decisions are what allow it to survive against large technology companies, Silicon Valley startups and model vendors.
8. The “Thin Middle Layer” Argument Misses That Engineering Abstraction Is the Most Expensive Tuition
张路宇 rejects the claim that “engineering is worthless”: “We believe engineering is the most valuable thing.” A platform must abstract the variable and invariant parts of its target users and business scenarios. Make Workflow nodes too granular and they approach programming, which ordinary users cannot handle; make them too coarse and the platform loses capability. There are countless such trade-offs.
He cites a report that roughly 80% of OpenAI’s new AgentKit code was written by Codex, and concludes that auto-generated code is not the same as accumulated engineering depth. The shape of a mature platform comes from years of working with customers, developers and production scenarios, backed by the “tuition” paid by countless teams putting AI into practice.
Dify’s key Workflow and RAG components can connect to existing knowledge bases and algorithms, while internally developed components can be hot-swapped and upgraded independently. Developers can move from 1.0 to 1.9 while preserving previously built suites; otherwise a single upgrade could bring the business down and wipe out its historical technology assets.
The engineering moat is therefore not code volume, but the sum of compatibility commitments, granularity judgment and real deployment experience. That knowledge is difficult to copy from a similar interface or a single code-generation exercise.
9. Staying in 0.x for Almost 2 Years Was Dify’s Restraint in Committing to Infrastructure
Dify remained on 0.x for nearly 2 years because the team believed the external environment was not yet mature. 张路宇 compared the situation to “sailing a ship through rough seas while trying to build a floating pier on top of it.”
The winners among model vendors, whether React would persist, whether fine-tuning would survive and how new modalities would emerge were all uncertain. Declaring 1.0 too early could cause major architectural tearing and drift, making the product untrustworthy to users. Entering 1.x formally means the team believes external variables have stabilized enough for the current architecture to be trusted for 3 to 5 years.
When 曲凯 asked about upgrade bugs reported by the community, 张路宇 did not evade the issue. The project is developed jointly by the core team and more than 1,000 community contributors, making quality coordination extremely difficult and requiring exceptionally strong automated testing. The community edition adopts new features such as MCP faster, while the enterprise edition ships more slowly with more testing, trading speed for lower risk.
10. Workflow Checkpoints Will Not Disappear; They Will Be Repartitioned as Models Improve
When enterprises hand production processes and reasoning logic to LLMs, they need reliability of 95% or higher. Take automatically finding a home, communicating and signing a lease: the task inherently contains multiple steps. Future nodes may become fewer and their granularity larger, but that does not mean the workflow disappears.
The extreme form of a single-node Agent is “input box—result”: if the user is dissatisfied, discard it and start over. As long as users need control, review or the ability to return the model to a human at a critical point, the workflow must expand. 曲凯 cites the repeated “draws” involved in generating images and video; 张路宇 points out that a person still makes the final selection, which is effectively human-supervised operation, while enterprises generally want unattended operation.
路宇 uses DeepSeek as an example, describing its technical direction as a gradual shift from today’s High Code—large-scale coding and extreme reliability—toward intelligence. The transition may take 3 or 5 years, but it can continue delivering reliable products along the way. Dify’s own route is to keep delivering production-ready technology: the technology changes every year, but the risk line does not. “Like the Toyota cars driving all over the streets outside, Toyotas that don’t break down,” users should always feel at ease, safe and confident.
Another route charges directly toward AGI and tries to solve every problem with a single bot. It may deliver spectacular demos, but it is also more likely to leave users frustrated, anxious and feeling “I’ve been fooled again.” 曲凯 warns that the pragmatic route will lose out in capital markets and social-media narratives; 张路宇 accepts that noise exists but will not change the risk line because of it.
11. Neuro-Symbolic AI Is the Theoretical Anchor Keeping Dify from Chasing a Pure-Model Route
张路宇 describes Dify’s direction as neuro-symbolic AI: neural networks such as Transformers handle association, generation and expansion, while early symbolic AI handles judgment, logic and causality. A truly effective Agent will ultimately need a combination of both.
His bio-inspired explanation is that the human brain operates at low energy not only because of neural association, but also because it contains fast classification mechanisms similar to binary trees: old person or child, boy or girl. Symbolic judgment compresses the search space, allowing people to make rapid predictions based on causal understanding.
Faced with the grand narrative that “pure neural networks can solve everything,” his response is: “I don’t believe this. I have never believed this.” A clear view of intelligence means the team does not need to change technical direction whenever a single impressive case appears.
12. Multi-Tool Combinations Will Persist; the AI Programming Paradigm Has Not Converged
It is not unusual for an enterprise to use Dify, LangChain and n8n at the same time. A new Agent programming paradigm has not yet formed, each platform solves local problems in its own way, and enterprise engineering capabilities vary. 张路宇 observes that North American programmers are somewhat stronger technically and tend to prefer LangChain, while some East Asian markets have greater need for low-code interfaces.
曲凯 proposes “fast-fashion SaaS”: will enterprises let finance, tax and business staff build Workflows themselves alongside technical teams instead of buying generic third-party SaaS? 张路宇 believes that is only half right—interactions and workflows can be customized, but stable data structures such as invoice fields, sensitive attributes and long-term storage cannot be changed arbitrarily.
Future programming may spend more effort defining outcomes and stabilizing structures, leaving AI to generate flexibly outside those structures. AI will reshape SaaS’s variable layer, but it will not eliminate data constraints and long-term consistency from software engineering.
13. Agents Advanced Faster Than Expected, and Dify Paid a Growth Price for Restraint
张路宇 admits that from around GPT-4 onward, it became difficult to predict when models would break through to a sufficiently high level of autonomy. Several model leaps pushed Agents ahead of expectations in areas such as programming. Companies like Cursor positioned themselves before the technology was mature and waited for the capability inflection point, roughly “waiting under the tree for a rabbit.”
For products seeking to capture a short-term trend, betting early on capabilities such as Nano Banana may be the right strategy: once a model becomes novel and interesting enough, the first representative product can absorb a concentrated wave of traffic. Dify’s risk appetite, however, means it will not trade production reliability for those odds.
The clearer mistake was in team building. At the start of 2024, the company had only a dozen or 20 people, while the business had grown severalfold by year-end without a corresponding talent reserve. When growth arrived overnight, it could not immediately hire enough excellent people, and the team struggled badly as a result. 张路宇’s postmortem is that “the conviction was not strong enough,” though 曲凯 also points out that preemptively expanding a domestically funded startup is itself a difficult choice.
14. Dify’s Mission Has Shifted Toward Rethinking Organizational Production Relations
张路宇 believes product development is still not fast enough. The root cause is not that the company failed to understand short-term competition early on, but that it failed to clarify its 5- to 10-year mission. The original goal was simply to build a popular, technically strong and well-liked open-source middleware product, without connecting it to changes in social structure.
He now defines the mission as “thinking about the future production and organizational relationship between people and AI.” When companies of a dozen people, a few people or even one person use AI to handle large-scale processes, the platform must adapt to entirely new forms of production and organization rather than merely helping individuals build a few small applications.
This also extends Dify’s early idea of “technological equality”: enabling people without years of technical training to automate and intelligentize their professional knowledge, becoming “10x engineers” or 10x professionals in other roles. 张路宇 says this no longer needs to be questioned and may become obvious within 2 to 3 years.
15. Enterprises Have Moved from Chatbots to Production Networks with Hundreds of Nodes
Two years ago, the common use cases were still companionship, knowledge-base Q&A and customer service. Even then, 张路宇 believed 20 chat windows would not last, because no one wanted to talk to them one by one every day. Bots were more valuable as a way to help users quickly understand AI and experience “the godlike feeling of a creator.”
Today, the most complex Workflow Dify has seen contains 400 or 500 nodes, connecting multiple internal organizational relationships and data sources before generating a production result. 张路宇 admits that even he could not have imagined this level of complexity early on.
He uses Anker as an example: it may have built more than 1,000 workflows with Dify, integrated tens of thousands of atomic capabilities and treated Agent resources as equivalent to HR headcount. A new business is first broken down into capabilities and SOPs, after which each step is assessed: “Do I put a person here, or an Agent?”
曲凯 is surprised that this stage arrived so quickly. 张路宇 stresses that it is not an isolated case; the spread of open-source, low-cost models driven by DeepSeek has further accelerated the entry of neutral middleware into production systems.
16. “Nonstandard” Is Dify’s Standard Use Case, and Its Form Looks More Like an Enterprise Operating System
Standard needs in finance, tax and marketing already have specialized vertical SaaS products, and a specific problem may cost only $20 to solve. Once a scenario becomes highly standardized, a specialist vendor can usually go deeper. Dify spent time considering whether to extract standard templates, but found that its best customers were instead all “nonstandard.”
These customers need glue to connect existing products, data, people and workflows into complex collaborative relationships; there is no uniform answer across different job functions and organizations. Processes that once required expensive IT integration can now be built at far lower cost with LLMs, AI coding and collaborative canvases.
张路宇 envisions not a simple scaffolding tool but a new form of organizational collaboration: internal atomic capabilities are inserted into a unified Hub, people and Agents collaborate on the same platform, and it is made explicit who designs the process and who handles human-feedback checkpoints.
In the future, employees may open a dashboard in the morning and see tasks waiting to be claimed, reviewed or given feedback. Entering a workflow reveals the full picture of business production. “You can think of it as an operating system”—its core objects are not computer resources but intelligent enterprise production.
17. Open Source Opened Japan; Global Organization and Context Capabilities Determine the Next Leg
Dify now has more than 110,000 GitHub stars and ranks among the top 50 open-source projects globally. 张路宇 says around 20% of Fortune 500 companies use it, while the company operates in what is essentially a “zero-sales” mode. He considers open source the most important decision: without it, Dify could not have solved global distribution, trust, technical standards and model neutrality at the same time.
The Japanese market may already be close to a “monopoly position” and has become a phenomenon: Dify appears on television, in bookstore textbooks and in coffee shops. It already had a Japanese version when the product launched in 2023. With a shortage of technical workers and highly process-driven businesses, local users saw Dify “the way an accountant who once only prepared financial statements saw Excel”; when the market exploded, the company had no employees in Japan, while the team now approaches 10 people.
Globalization was designed around asynchronous collaboration from day one: hiring people with international backgrounds, using US SaaS products to connect email, finance and other tools, and learning local product ecosystems along the way. Internally, Dify uses “trust first” systems such as giving investors access to all documents, using the same data internally and externally, and issuing new employees a $1,000 credit card subject to review afterward, following Drucker’s principle of “eliciting goodwill.”
张路宇 believes models are already “good enough, even luxurious” for most people: confidence in solving problems has risen from 60%—70% 2 years ago to above 95%, making context the bottleneck. A model is like a giant ball of data; the problem is entering the correct “password.” Dify needs to provide a map for finding answers, while organizations need shared Memory and an information bus, perhaps even a “chief context officer.”
18. The Final Moats Are Exceptional Individuals, Organizational Learning and the Founder’s Long-Term State
张路宇 does not predict model capabilities: “I’m the person teaching everyone how to use electricity. Why would I study how to generate it?” But as technical variables stabilize, Dify has extended product planning from 6 months to 3 years, focusing on Memory, MCP, connections to the real world and ecosystem collaboration.
曲凯 cites an MIT report saying that 95% of enterprise AI pilots fail. 路宇 summarizes the problem as insufficient tools and insufficient organizational learning capability, and believes the biggest opportunity over the next few years will be “building bridges”: infrastructure, human-computer interaction paradigms, reliable Workflows and the broad dissemination of knowledge about how to use AI effectively.
Some of the people who use models best are precisely the ones who say they are “not enough”—the models are not fast enough and do not offer enough Tokens. The team even tracks each developer’s model consumption. The capability gap may be more extreme than the 80/20 rule suggests, and these people share one trait: curiosity, “constantly walking along the edge of the model’s capabilities.”
张路宇 rejects the romantic explanation that Dify won after being underestimated through “careful cultivation and deep accumulation.” He prefers to say that the team applied methods from mature markets, consistently fixing labels such as engineering-first, stable and reliable, open and collaborative. The founder must also stay actively happy so ideas can emerge naturally: he rejects long-term 996, but also opposes confining creativity to a rigid 8-hour day. The optimal state is entering a flow where time no longer matters once everyone is aligned on the goal.