151: MuleRun’s Chen Yusen on Claude Code and Disposable Software
Summary
MuleRun’s core bet is not to build another coding agent, but to treat Claude Code as the strongest general-purpose agent available and build creation, execution, and transaction layers around it. Chen Yusen sees coding as the bridge through which AI operates the computer world: as long as a task does not blow up the context, sufficiently complex prompts, skills, and runtimes can already support complex work. The emphasis will shift from simply writing code to “choosing tasks with taste, elegant ways to solve them, and what kinds of problems to solve.”
MuleRun moved from Web coding to an Agent Marketplace, then from a shelf-style marketplace to an Agent Builder and conversational entry point. The first pivot was driven by insufficiently rich supply: the team’s early Web coding product produced results similar to Lovable and v0.dev under the same prompt. Adding frameworks such as n8n and LangGraph did not produce the expected volume of PMF. The second adjustment reflected the fact that a shelf made it difficult for new users to understand and choose. The new approach uses Claude Code directly as the base agent, letting nontechnical users define tasks in natural language and select skills and runtimes. “Even low-code is somewhat complicated”; true democratization means not having to understand parameters, loops, or conditions.
As software development costs fall, the market unit will shrink from standardized products serving 1,000 or 10,000 people to long-tail tools serving 1 to 10 people—eventually reaching “disposable software.” Chen Yusen imagines code being “precisely created, executed, completed, and destroyed” for a single purpose, much like on-demand manufacturing with 3D printing. But AI is also manufacturing: every production run consumes tokens, compute, and electricity. As long as one model remains materially ahead, costs will not fall to negligible levels through price competition in the short term.
The Marketplace may evolve from display and recommendation toward conversation, dispatching the relevant supply when users express a need; the mechanisms for transactions and allocation remain unsettled. MuleRun may already have 100 or 200 good agents, but new users still face a blank shelf, try two that do not work, and leave. The front end may eventually look like a general agent, while the back end matches tasks against 1 million or 10 million prebuilt offerings. The unresolved question is revenue sharing: moving from the clear price of “one product” to how creators are rewarded when their agents are called, and whether the market should center on sharing or transactions.
What the platform must do is more than “wrap” an existing model: it must make long-tail agents simple, reliable, measurable, and settleable. MuleRun wants an agent to fail at most once in 100 executions for a given task, and once in 1,000 for some tasks—not “fail five times out of ten.” That requires the unglamorous work of skills evaluation and security audits, continuous benchmarks, runtimes, credential permissions, reviews, refunds, and large-scale concurrency. Chen Yusen’s engineering principle is “lots of SOPs plus a small amount of large-model intelligence”: deterministic work goes to code, while the model handles judgments about the real world that are difficult to express as if-else statements.
Anthropic launching Claude CoWork itself shows that more people are realizing a coding agent can become a general-purpose agent, while also compressing startups’ time window. CoWork mainly packages capabilities already available in the terminal for ordinary users who will never open one. MuleRun is betting on controlled cloud accounts, preinstalled tools, a skills marketplace, deployment, and sharing, with a stronger focus on work and prosumers. Chen Yusen admits the team “moved too slowly,” because the opportunity lies just before non-consensus becomes consensus, and ecosystem products are often built by launching an “80%-finished product” and iterating with users.
Small but real examples of willingness to pay have already emerged on the demand side, showing that a translation layer between AI capabilities and ordinary users can charge for access. When Nano Banana was popular, one creator packaged a public prompt, Google calls, and an n8n workflow into an agent charging a few cents per use. Chen Yusen recalls that it may have generated about $1,200 in 3 days, while stressing that he does not remember the number exactly. Earlier, people used Midjourney to make pet portraits for RMB10 each. The “AI lackey” makes money not because the model lacks the capability, but because access, search, prompting, and operational barriers still keep much of the demand out.
For an entrepreneur who has lived through the rise and fall of 4 companies, MuleRun is also a restart—from technical arrogance back to “putting your heart into it.” Changting Technology grew from its founding in 2014 to its sale to Alibaba Cloud in 2019, with revenue at least doubling almost every year. His subsequent parallel ventures in gaming and data security left him suffering severe insomnia and self-doubt in the second half of 2023. His constraint now is to choose work matched to his abilities and understanding, and to get directly into the arena. Over the next 6 months, he will first see whether more people without coding skills can create useful agents. “There is no ‘good enough’ in entrepreneurship; only by continually pushing yourself can you find even a sliver of a chance.”
Deep dive
1. Web coding proved the bridge value of code, while exposing the ceiling of homogeneity
Chen Yusen started from Cursor finding PMF in late 2024: programming is an extremely general capability, and once models can replace or assist humans in programming, AI gains a bridge to completing more computer-based tasks.
In April of the following year, the MuleRun team built a Web coding product similar to Lovable and v0.dev. Under the same prompt, the webpages and results for most tasks were similar across several products. “It didn’t seem like we had done it better than others,” because the ceiling was largely determined by the Claude API that everyone was calling.
His observation was that this race had not produced a meaningful reshuffling of the rankings through product design. Model upgrades made all products stronger together, but did not automatically bring later entrants traffic or differentiation. The team therefore abandoned simply copying Web coding and began looking for the platform opportunity created by an explosion in supply.
2. Claude Code turned the coding agent into a general-purpose agent
Chen Yusen’s clear judgment is: “Actually, the strongest general-purpose agent right now is Claude Code.” In his view, anyone who recognizes this will naturally ask the next question: if it can write code, why can’t it use code, tools, and interfaces to do other computer work?
Anthropic employees took Claude Code, originally built for coding, and used it to read email and handle miscellaneous tasks. That gave him a crucial insight: users never wanted merely to write code; they wanted tasks completed. Coding was simply the way the agent temporarily constructed tools.
His strong hypothesis comes with a time condition: “Once large models have coding capabilities, in theory they can do everything humans can do on a computer within 1 year.” What is currently missing may not be intelligence, but task knowledge, suitable environments, and reliable ways to deliver results.
3. Complex prompts can already carry complex work, but uncontrolled context causes collapse
Chen Yusen found that current models can indeed handle long stretches of complex work when given sufficiently complex tasks and prompts. The condition is that the context cannot “blow up”; once it does, hallucinations multiply and the task goes out of control.
As a result, the real engineering work is shifting toward context engineering: when to give the model which materials, how to write prompts, how to compress or load information in layers, and which steps should never enter the context at all.
This also changes how agents are built. In the past, frameworks such as LangGraph were commonly used to write SOPs and models into code. Now a strong base agent can understand the task, then dynamically write and execute the code needed to complete it.
4. Human brains and offline knowledge form the original supply of the Agent Marketplace
MuleRun’s underlying judgment is that substantial professional knowledge still resides in people’s heads, offline processes, and organizations, and has not all entered models. Reality also continues to generate tasks that did not previously exist.
If domain experts clearly explain their experience, judgment criteria, and work steps, Chen Yusen believes it may be possible to “copy 80% or 90%” of an individual’s capabilities into an agent. Others facing the same problem would no longer need to operate a computer from scratch, and could instead pay an affordable fee to reuse that capability.
This creates the platform logic: experts package tacit knowledge into supply and earn money, demand-side users “solve big problems with small amounts of money,” and the platform makes agents discoverable, reliable, and repeatedly transactable.
5. Low-code still keeps true domain experts outside the gate
The team initially supported developers uploading applications and made room for supply such as n8n workflows and LangGraph agents. It soon found that people who could use low-code proficiently could often already code, and were not the audience the platform most wanted to activate.
Chen Yusen’s threshold is deliberately aggressive: “You shouldn’t need to know what a parameter is, what drag-and-drop is, what a loop is, or what a condition is.” Ordinary people should only need to describe the needs, processes, and judgments they already handle.
An HR recruiter knows how to screen candidates every day. A customer-service representative knows which systems to search for knowledge and how to handle complaints, returns, and exchanges. What they lack is not business knowledge, but the tools to turn that knowledge into executable software.
6. Large models supply the judgment glue that workflows cannot encode as rules
Traditional CRM, ERP, and IM systems have long existed in the workplace, but specific actions still depend on individual judgment: whether to replenish insufficient inventory, which process a customer request should enter, or which candidate meets an implicit standard. These are difficult to enumerate as rules.
Chen Yusen believes large models are suited to replacing these judgments that pose “not that great an intellectual challenge” but cannot be fully encoded. Standard steps should continue to be executed by software, with the model interpreting business context in between. Only then can end-to-end automation work.
The gap between semi-automation and full automation is therefore amplified. AI can work 24 hours a day and create 100 or 1,000 copies running simultaneously. Once accuracy is high enough, the efficiency gain is no longer just a matter of saving a few clicks.
7. Runtimes and skills determine whether Claude Code can truly operate a computer
Without a browser, Claude Code cannot perform operations inside a browser. Giving it a Windows machine does not mean it inherently knows how to play a particular game. The model needs an operable environment, as well as task rules and domain knowledge.
MuleRun wants to put browsers, Excel, designated SaaS products, tools, and the relevant permissions into a controlled runtime, then encode working methods in skills. Creators mainly describe the need; the system combines the base agent, skills, and a container or virtual machine.
Chen Yusen’s extrapolation is that with “sufficiently rich runtime environments and sufficiently clear skills,” Claude Code could theoretically complete any computer operation. MuleRun’s job is to turn that capability into a product that can be delivered repeatedly.
8. Good prompts already have commercial value; most of that value simply has not been captured by a transaction system
Chen Yusen points out that excellent prompts are often discovered through the first creator’s effort, but usually spread for free as social content. Nano Banana’s desktop figurine trend is a representative example; it is now difficult to trace who originated it.
Actually using the capability still involves several layers of friction: accessing Gemini or Google AI Studio, finding the prompt on Twitter or elsewhere, uploading the correct image, and completing the call. For people outside the tech community, “using a search engine well” is itself a skill.
One creator on MuleRun packaged the process into a simple n8n agent and charged a few cents per use. Chen Yusen recalls that it may have earned about $1,200 in 3 days, but specifically says he does not remember the number. The signal is not the precise revenue, but that convenient packaging clearly generated usage.
Cheng Manqi added that an 18- or 19-year-old restaurant server had also joined a group about “making money with AI,” handling profile pictures for other people. Chen Yusen calls such a person “an AI lackey”: the model already knows how to do the work, but people who understand access, prompts, and operations can still charge for their translation ability.
9. n8n, Dify, and Coze occupy adjacent positions, but their priorities are completely different
Chen Yusen believes early workflow markets all competed around creation tools. n8n had a marketplace very early, but it mainly traded source-code templates and failed to capture the transaction value of repeatedly invoking the same workflow.
Dify’s commercialization is more To B-oriented, supporting internal enterprise workflows in Japan and the US. Because enterprise systems cannot be exposed externally, Dify’s value is deploying processes inside the internal environment; security is the first constraint.
Coze’s early output consisted largely of chatbots, and it long bet on drag-and-drop low-code. It later began moving toward natural-language creation similar to Web coding. Chen Yusen is more positive about that change, but believes ByteDance tends to build closed-loop systems and will likely enter the marketplace eventually.
These products could be upstream creation frameworks for MuleRun, or competitors. “Who doesn’t want an App Store for a new era?” The difference is not whether they want to build one, but whether they enter through tools, enterprises, content, or transactions.
10. The same democratization of creation can grow into a social network, content platform, or productivity marketplace
Chen Yusen characterizes Ant Lingguang as more “make for yourself, use for yourself,” and sees YouWare as “make for others to use.” Cheng Manqi adds that once something is created, use by oneself and by others naturally coexist. What the platform determines is which behavior gets priority and how it is rewarded.
Chen Yusen uses smartphone photography as an analogy: after photography was democratized, the same photo format gave rise to different products such as Instagram and Snap. Web coding’s output is similarly new, and companies are deciding whether it is content, a work, social expression, or a productivity tool for solving workplace problems.
Games, galleries, and entertainment works on YouWare look more like content. MuleRun places greater emphasis on using outputs to solve workplace problems. Its ceiling is not another asset community, but “the world’s largest labor-outsourcing company.”
11. The shelf marketplace produced no expected PMF after 3 months, pointing first to supply
MuleRun initially assumed that every highly specific scenario might support roughly 100 DAUs: users would return every day and have a specialized agent solve a repetitive problem. After about 3 months online, the team saw only a few or perhaps 10 strong daily users in a small number of scenarios.
The growth team brought in many registrations, and the platform completed its transaction functionality and listed a batch of supply. But Chen Yusen admits they “didn’t see enough PMF happening at scale.” The team was forced to distinguish between insufficient users and a shelf that simply lacked enough rich, reliable offerings.
The classic cold start of a two-sided market was especially acute here: without users, creators asked why they should upload; without supply, users asked why they should come. Chen Yusen calls it the difficult-to-achieve “meeting the middle.”
Existing creation, deployment, manual review, and multi-framework compatibility were all too complex. The core long-tail hypothesis could only be tested if the supply process became “convenient enough and short enough,” making Agent Builder a prerequisite rather than an add-on.
12. The platform cannot rely mainly on official supply to start the ecosystem, but must personally defend quality
Chen Yusen initially insisted that the official team should not heavily enter the market, reasoning that if Apple filled the App Store with its own apps, the ecosystem would be difficult to establish. MuleRun provided only foundational capabilities such as General Browser Operator.
Cheng Manqi asked why the team did not anonymously create a batch of agents to heat up the market first. Chen Yusen’s practical answer was simple: it was not that they did not want to; a team of about 50 people also had to build deployment, transactions, and framework-neutral infrastructure, and “couldn’t do it all.”
Looking back at the 3 months, he admits that if the team had built more well-chosen scenarios early on, “it would not necessarily have been a bad thing.” If external creators did not quickly embrace the new paradigm, the team could also use Claude Code and skills to produce a large amount of relatively useful supply within 1 month.
Regardless of who creates it, listing standards must be strict: a given task may fail once in 100 executions, or once in 1,000 for some tasks. If it “fails five times out of ten,” it is garbage that should be removed.
13. Code handles determinism, large models handle judgment; engineering competition depends on how quickly it can be shipped
Chen Yusen has repeatedly read The Bitter Lesson. Its core insight is consistent with Claude Code’s code execution: anything that can be solved reliably with code should not be stuffed into a long context and left for the model to calculate.
“Which is larger, 9.11 or 9.2?” can test a model, but in a production system the answer should simply come from code. The model’s job is to identify the problem, generate or select the correct program, and read the result—not replace deterministic computation with expensive and unstable reasoning.
MuleRun therefore bets that “lots of SOPs plus a small amount of large-model intelligence” will remain effective unless a powerful AGI capable of any task appears directly. If that day comes, it will not only disrupt MuleRun, but almost every company.
Until then, the team must handle startup speed, concurrency stability, image configuration, and environment selection. Not letting users choose skills is the most natural experience, but may require training a small model. There is no cost-free answer across accuracy, experience, and time to market.
14. An agent platform may absorb vertical supply, but will not immediately replace a memory-enabled general agent
Chen Yusen uses Chinese e-commerce as an analogy: independent sites struggle to acquire traffic cheaper than platforms, and small consumer brands typically rely on Taobao. Likewise, many vertical AI products for recruiting, marketing, image creation, and other use cases may be better off growing inside a unified platform.
This means a platform could become larger than the vast majority of standalone agents, but he does not view general-purpose agents as competitors on the same level. The user habits differ, and general-purpose agents also possess long-term memory.
He uses his own difficulty leaving ChatGPT as an example: after years of asking questions, the product broadly understands what kind of person he is, and new questions can draw on past memory. MuleRun is more like a market for precisely calling specialized labor and does not need to replicate that relationship.
If supply reaches 1 million or 10 million agents, MuleRun may grow an entry-point agent: users state only the problem, and the entry point finds the most precise prebuilt supply. The platform could eventually become something close to general-purpose capability, but the path would come through the ecosystem rather than a single model.
15. Skills are fundamentally about layering context on demand, not stuffing every instruction into the model at once
Long prompts often lose information through context compression on complex tasks. Skills initially expose only meta information about “what problem they solve”; once the agent determines that a skill is relevant, it reads the full content.
Most skills take the form of documents and prompts, but they can also carry programs, tools, or APIs. A skill can tell the model not only how to do something, but also “not to do it yourself,” and instead call existing code to obtain the result.
Chen Yusen sees skills as part of context engineering: the longer the context window, the more hallucinations increase and the lower the model’s “IQ.” Information that can be executed, retrieved, or loaded later outside the context should not remain there permanently.
This is also a key change from traditional software. When real-world problems cannot be perfectly classified, branches are no longer hard-coded through if-else statements. The model observes the task and skills simultaneously and decides which path to enter.
16. The value of a Skill Marketplace comes from evaluation, security, and tool coordination
There may already be tens of thousands of skills on GitHub, with 3, 5, or even 10 versions for the same task. MuleRun has to answer not whether a skill exists, but which one performs best on comparable tasks.
The team is building skills evaluation: compare completion performance on the same task, then let creators call the better-performing candidates with one click. Frequently used skills may be written officially, or optimized jointly with strong developers.
Security audits are equally indispensable. A skill may contain malicious behavior. If Claude Code runs on a user’s computer, it could even reverse-shell the machine and give an external party control.
Browser automation is a priority because a large amount of work already takes place in browsers. An actually effective browser skill often comes with tools, so skills, tools, and runtimes must be optimized together rather than placed in 3 separate directories.
17. As development costs approach zero, software will move from industrial products to “disposable goods”
Chen Yusen uses a colleague’s 3D-printing analogy to explain the long tail: industrial production can serve only enough people, while 3D printing can make something for oneself and 3 or 5 friends—an item with little industrial value but considerable personal need.
Traditional software needs to cover the shared needs of at least 1,000 or 10,000 people to spread development costs. Web coding and powerful agents shift organizations from centralized development to distributed creation, making needs of 1, 2, 3, 5, or fewer than 10 people worth serving.
The more distant end state is that “software in the future will actually be disposable after a day.” Code will no longer exist as a long-term product, but will be precisely created, executed, completed, and destroyed for one purpose.
Chen Yusen also acknowledges that the industry cannot jump directly to the end state. The more realistic short-term value is to build a batch of agents with few users but that a small number of people use “with great satisfaction” every day, while letting creators produce once and deliver repeatedly.
18. AI is manufacturing; tokens, compute, and electricity do not disappear because software becomes temporary
Every 3D-printed item consumes filament, and every AI task consumes tokens, compute, and energy. So while “disposable software” serves the long tail, it is not a zero-marginal-cost software utopia.
Chen Yusen believes tokens are unlikely to become cheap enough to be insignificant in the short term. A price war will emerge only when 3 to 5 suppliers all have sufficiently good intelligence and no clear gap between them. As long as one provider is materially ahead, its leading capability retains pricing power.
The costs of the previous generation of models may eventually fall to reasonable levels, but demand for the strongest intelligence is nearly unlimited because AI is reshaping social labor and productivity itself. Once “the work on Earth is done,” humans will go to space to look for new tasks.
He views rising CPU, memory, and hard-drive prices, along with electricity shortages, as part of the same supply-and-demand chain, extending the logic to controllable nuclear fusion, Dyson spheres, and space data centers: new technologies create new demand for the next layer of technology.
19. Agent Builder stands on a base agent, saving the time required to rebuild an agent from the ground up
Chen Yusen breaks an agent into several core components: the agent loop that repeatedly interacts with the model and processes results; context engineering that controls context length and information flow; and tools that let the model operate a browser, create images, or call external systems.
General-purpose agents such as Manus design their own loops and context engineering. Frameworks such as LangChain and LangGraph provide adjustable implementations. Claude Code has already built both into a mature base agent.
As a company that pivoted again in January 2026, MuleRun did not have time to rebuild from the bottom up. It therefore chose to “wrap more deeply”: use Claude Code as the core, then add skills, runtimes, cloud editing, and deployment.
Users define a task and select the skills and environment they need to build their own agent. The underlying model could eventually be replaced by OpenCode. Once the team has more capacity, it may also move down into the loop and context-engineering layers, but the fastest path is to use the best base agent available today.
20. Reachability often drives mainstream adoption more than technical leadership itself
Chen Yusen admits that from a purely technical perspective, the additional layers MuleRun currently adds may “seem like they haven’t done anything.” But people from technical backgrounds easily underestimate the value of lowering the barrier, ranking it fourth or fifth and effectively deciding that it does not matter.
DeepSeek changed his thinking. o1 had similar capabilities at the time, but DeepSeek’s free availability, complete display of CoT, viral momentum, and playful Chinese phrasing made it easier for ordinary users to reach. The episode mentioned that it trained such a model using “only $5.57M.”
He similarly believes ChatGPT saw a major growth event in March or April 2025 because of Ghibli-style image generation, while Google’s growth owed more to NotebookLM than merely to improved Gemini capabilities. Technical advancement matters, but product wins are not determined by benchmarks alone.
Chen Yusen cites a Gemini 3 core engineer: scaling is nowhere near hitting a wall. Even if model intelligence stopped improving today, humans might still be 1 to 2 years away from exhausting the capabilities already available. That leaves ample room for application growth in 2026.
21. Claude CoWork validated the direction, but local desktops and cloud services serve different scenarios
Anthropic launched Claude CoWork during the same week MuleRun was internally testing Agent Builder, showing the team that “coding agents are general-purpose agents” was moving from non-consensus to consensus. Chen Yusen’s reflection is: “We spent slightly too long insisting on and wrestling with the marketplace.”
Claude’s desktop product started with Chat, then added Code, and then CoWork. Chen Yusen believes most of what CoWork does could previously have been done by telling Claude Code in the terminal. The core change is making the capability available to users who “will never open a terminal in their lives.”
CoWork can read local files, calendars, and login credentials, making it suitable for personal context. MuleRun runs in the cloud, allowing users to create separate accounts for agents with tightly controlled permissions, with a stronger focus on prosumers, work, and productivity.
MuleRun also tries to package deployment, tool downloads, environment configuration, and sharing together, so users do not hit one blocking point in an n8n workflow and become unable to continue. The team does not rule out building for the local device later, but the current competitive positions do not fully overlap.
22. The shelf will retreat to a secondary entry point, while conversation dispatches the ecosystem when demand arises
The old Marketplace may already have 100 or 200 decent agents, but new users do not know what the platform can do. If they click on 2 at random and neither solves the problem, they will quickly leave.
The discussion of application entry points in the episode summarizes 3 generations: display, recommendation, and then conversation. Chen Yusen believes that in the future, “I won’t recommend anything either; I can help solve some problems simply by talking with you.”
The new front end may look like a general-purpose agent: it completes whatever it can directly, and when it encounters a complex or long-tail task, it triggers an agent refined by a creator. Users do not need to search for a product or rewrite a complex prompt.
The problem shifts from discovery to revenue sharing. Shelf products can carry a direct price; the platform has not worked out how to reward contributors for calls made through a conversation, nor whether the market should center on sharing or transactions.
23. “Using agents to build agents” will become consensus; the winner will be whoever activates expert supply first
Dify and Coze, which was folded into the Volcano Engine system, are more To B-oriented; Coze Space, AnyGen, and Manus look more like general-purpose agents. Chen Yusen believes a general-purpose agent can absolutely develop the ability to “manufacture other agents,” and Claude CoWork can also be understood as a product produced by Claude Code. He is unsure whether Manus will do the same.
Claude Code’s “one version every day” showed him a new iteration paradigm: once productivity is unleashed, even global software can update at high frequency. Ecosystem products that wait until they are fully mature to launch may lose to teams that release an 80%-finished product first and evolve with users.
He is skeptical of the old path of letting users make only simple webpages and games, arguing that such content has “thin value.” More valuable is enabling industry experts to turn offline SOPs and domain knowledge into agents.
For 2026, he hopes to see AI-native applications with annual revenue of $500M or even $1B, rather than only the chatbots everyone already uses. For MuleRun, the most meaningful model progress would not be higher scores in math competitions, but more accurate and efficient tool use.
24. Changting’s success came from matching capabilities, but also planted the illusion of being able to do anything
During university, Chen Yusen shifted from gaming to cybersecurity and formed the Blue Lotus team with partners. They won almost every domestic competition and generally achieved the best results among Chinese teams overseas. The DEF CON final required 40 to 50 hours of continuous competition; the 8 people on site attacked other servers, patched their own services, and maintained health checks at the same time.
From top competitors such as CMU PPP, Ricky, and George Hotz, he felt that “the gap between geniuses and non-geniuses in the computer industry is boundless.” Continuous positive feedback pushed this group toward entrepreneurial directions that genuinely matched their strengths.
Changting Technology was founded in 2014 and sold to Alibaba Cloud in October 2019. Chen Yusen spent nearly 7 years on his first venture. Company revenue at least doubled almost every year. He started as the first salesperson; despite rejection, the monthly chart kept moving upward in retrospect.
The decision to sell was driven by cloud providers absorbing infrastructure demand and the financing environment beginning to deteriorate in 2018. He later concluded that the direction was right but the timing was slightly early: the financing market peaked only in 2021, when he happened to be building another security company, showing that identifying a trend correctly does not mean timing the trade precisely.
25. The second venture lost focus across 2 fronts, let capital inflate its ambitions, and lacked humility toward the industry
After leaving a major technology company in 2021, Chen Yusen founded a gaming company and a data-security company at the same time. He felt “capable of anything,” had the credibility of a first exit, and could raise substantial capital, so he believed both could be done well together.
The gaming team originally wanted to make a small-team PC game with innovative gameplay and themes. The industry’s classic warning is that the first 10 games are highly likely to be garbage, but the team was overly confident in its first product. Excess capital also pushed them toward a project large enough to match the financing scale, rather than spending a small amount to fail first.
He now admits that he did not show enough humility toward an unfamiliar industry: “Don’t think as if everyone who came before you in the industry simply didn’t know how to do their job.” Learning quickly does not mean getting it right the first time, and past success can magnify a bad bet.
The data-security company continued to operate, built products, and served good customers, but commercialization in Chinese enterprise software and security has been difficult in recent years. Even with more experience and resources than the first time, rebuilding the same business after its window has passed remains extremely challenging.
26. The failure of 2023 took him from peak confidence into severe self-doubt
The low point came in the middle and latter part of 2023. Chen Yusen described himself as suffering from “severe anxiety—so bad that I couldn’t even get out of bed,” accompanied by serious insomnia and repeated questions about what had gone wrong and how he could have made so many bad judgments.
As the team declined, there were fewer and fewer happy moments. Internal conflicts and a loss of cohesion emerged together, and he could feel core members’ trust, recognition, and expectations of him gradually falling. He began reinterpreting his earlier success as, “Was it all just luck?”
The company still had tens of millions of RMB on its books. In theory, it could have shut down the old businesses and started over in AI, but his mental energy was out of sync with the external wave. He judged that he was not in the right state and that the probability of turning things around by spending the remaining cash was extremely low, so he shut down the gaming company: “If it failed, it failed.”
What actually helped was not hearing “success and failure are common in war,” but having friends accompany him to an internet café to play games for 2 days. He later returned to Alibaba to work on internationalization, reconnected with Silicon Valley and overseas AI developments, and began looking for a team and new direction only after recovering.
27. Starting over comes down to 2 constraints: choose work that matches your abilities, and genuinely put your heart into it
Chen Yusen’s first rule is not to let financing size, shareholder returns, or other external conditions push him to expand his choices. He should first do work matched to his abilities and understanding and that he genuinely believes in. “When you make choices sincerely, everyone accepts the outcome even if you lose.”
The second is to get directly into the arena: entrepreneurial experience cannot justify hiring 2 or 3 capable people to experience the wave on his behalf. Hiring, product, and customers all need to be tested on the front line; for second-time entrepreneurs, insufficient time commitment remains a common cause of failure.
He ultimately condenses his values into “putting your heart into it.” In his first venture, he was awake thinking about how to make it succeed; in the second, pride and “that’s probably good enough” entered the equation. But “there is no ‘good enough’ in entrepreneurship.” Carefully completing a task is not the same as genuinely caring, thinking deeply, and creating differentiation.
At 33, he admits to having less energy than at 22, but more experience and a mindset rebuilt after recovery. MuleRun’s next milestone is to build a sufficiently simple creator tool within 6 months and activate nontechnical people who have never imagined they could make software. True success should exceed the team’s expectations, just as the App Store grew The Talking Tom and Angry Birds from creators’ unexpected answers.