Original · Unique Research · 2026-07-24
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, five insight sections, speaker list, and the full roundtable transcript. Business, project, and revenue figures are speaker self-reports attributed to the named individuals, not independently audited findings. Product names, company names, and named people are preserved as source attributions.
AI Industry Observation
Products Can Be Copied; Judgment Cannot
"If you build a large model project like traditional software development, the project is doomed to fail."
The person who said this is Chai Yatuan (柴亚团), founder of Rongzhi Information (容智信息). His company specializes in Agent implementation for government and state-owned enterprises. Last year they delivered 20 projects, ranging from 300,000 RMB for small ones to 4 to 5 million for large ones. By all rights, he should be the last person wanting to offend clients. But he insists: clients demanding AI be 100% accurate with predictable results have this fundamentally wrong from the start.
I heard this at a roundtable. The panel had five companies: ToB implementation, vibe coding, no-code, personal Agents, and zero-code content platforms.
Don't Celebrate the Demo in 15 Minutes Too Soon
Chai Yatuan, co-founder and CEO of Rongzhi Information, laid out a blunt enterprise standard: an AI project must be acceptable — the result must at least exceed human performance.
If a person handles 100 orders a day, the Agent must handle at least 1,000 to 2,000. If a person achieves 85% completeness or accuracy, the system must cross that line too.
Rongzhi landed about 20 projects last year, ranging from about 300,000 to 4 to 5 million RMB each, serving large clients like government and state-owned enterprises. At this scale, building an interface that "looks like it runs" is meaningless — business results must actually deliver.
But it's not that simple. Many clients initially evaluate AI like traditional software, expecting input A to always produce A. Chai Yatuan's judgment is that this drains AI's value. Models have creativity and instability; the key is to first define the range within which they can operate, then clearly state the taboos and constraints.
"The passing line for an AI project is that it must run faster than humans AND allow results to fall within an understood, accepted range. A model can fill capability gaps, but it cannot fill gaps that management hasn't figured out."
Between "Software Can Be Generated" and "Software Can Be Run," There's the Final 10% of Wrapping Up
Guo Yu (郭宇), co-founder of superun.ai, faces a different group: business owners, top business leaders, and entrepreneurs. They understand their own business but have rarely done design, product, or technical management.
In the past, this type of need had only three paths: hire an outsourcing shop, wait for the internal IT team, or buy a standard SaaS product. The first two are expensive or slow; the third often can't accommodate an enterprise's personalized processes. Now, users can collaboratively build a requirements document through chat, choose a design, and ask the AI assistant along the way how to decompose features and how to advance the release.
Guo Yu says they have already delivered to thousands of entrepreneurs. For these users, true delivery doesn't stop at generating a page — it's a product that can go live, be used, and be continuously iterated.
"Making a product is easy. Making a product with taste, stability, and long-term usability is hard. AI output 90% of the time stays at Demo level. The real final 10% — concurrency, security, launch, and payments — all come up."
A Black Box You Can't Understand Will Eventually Become the Product Owner's Problem
So if AI handles all the processes, does the user just need to say one sentence?
Jiang Yaokai (蒋耀锴), founder and CEO of Hanzi Tech (函子科技), gave a more sober answer. He attributes the value of no-code to "visibility": logic is presented visually, so even without coding knowledge, the product owner can understand what the system is doing.
His concern is concrete. If AI-generated code becomes a black box, what assumptions, dependencies, and limits exist in the system will, over time, sink into compressed context. People with computer training can trace the code; business owners cannot, making verification impossible.
"Vibe No Coding must let product owners understand the process, find the assumptions, and participate in verification, even if they've never studied code. Whether users can run a business on it is the result."
Hanzi's standard is simple: users must make money on it. A PE fund manager who loves sports cards first built a card collector on the platform. As users grew, consignment trading, group buying, logistics automation, and grading gradually emerged. Jiang Yaokai says this person still operates alone and did 6 million RMB in revenue last year.
When Coding Rights Go to Ordinary People, Long-Tail Needs Finally Surface
In this discussion, Lingzhu AI (灵珠 AI) CMO Wu Di (吴荻) pushed the perspective further. She says Lingzhu's target users are 1 billion ordinary people; among public beta testers, the 30–50 age group accounts for 60%, and children aged 8–12 are the second-largest group.
This user structure is interesting. Coding tools originally attracted programmers, but zero-barrier creation is reaching doctors, lawyers, small business owners, and children. Someone used it for interactive medical science; a parent turned a mistake notebook into an interactive mini-program. A sophomore student validated a card game online and then made it into a physical product, with orders approaching 500,000 RMB.
These may not fit traditional SaaS standardization logic, but they genuinely respond to small human needs. Wu Di notes that product managers used to routinely cut long-tail needs — the reason isn't complex: customization is expensive and products are hard to scale. After AI lowers creation costs, those previously cut needs have their first chance to grow on their own.
What Wu Di calls "coding rights" here goes beyond who can build management systems. Interactive games, knowledge popularization, psychological assessments, and all kinds of unnamed small works can also grow from one person's idea into something others are willing to play, use, and pay for.
Good Product Technology Should Disappear into the User Experience
When the conversation turned to the future, several technical terms appeared on stage. Jiang Yaokai talked about the Agent loop: letting the Agent continuously see errors, dependencies, and front-end rendering results during construction, then proceeding accordingly. His judgment is direct: whether human or AI, learning is fastest when receiving timely, rich, and targeted feedback.
Chai Yatuan hopes model capabilities continue to strengthen, making today's heavy engineering patching work lighter.
Guo Yu cares whether AI can continue to participate in product operations, growth, and iteration.
Ma Liang (马亮) sees foundational models and Harness frameworks mutually accelerating under real user data.
Wu Di pulls the question back to users. Ordinary people shouldn't first research which model to choose, nor should they need to understand backend skill templates. Technology selection, model adaptation, and process scheduling stay in the background; users just need to explain things clearly through conversation.
"One sentence from the user, Lingzhu runs itself ragged. A mature AI creation tool should have an increasingly simple front end, while the back end takes on more and more complexity for the user. When users no longer worry about code, models, and deployment, software truly transforms from a specialized skill for the few into a daily capability for the many to run their ideas."
More Conversation Details
Speakers
Chai Yatuan, Co-founder & CEO, Rongzhi Information
Guo Yu, Co-founder, superun.ai
Jiang Yaokai, Founder & CEO, Hanzi Tech
Ma Liang, CEO, Silicon Geek (硅基极客)
Wu Di, CMO, Lingzhu AI (灵珠AI)
Host
Sang Zhuohao, Senior Director & AI Lead, Focus Media (分众传媒)
Who Do You Serve and What Pain Points Do You Solve?
Sang Zhuohao: Let me throw out the first question. You should all talk about what group you serve today, what their biggest pain points are, and why your product satisfies those pain points so well.
Chai Yatuan: Hello everyone, I'm Chai Yatuan. Many people know Rongzhi Information from our early RPA work, but starting in 2023 we launched a systematic transformation. After nearly four years of deep work, we've completely transformed into an AI Agent company focused on ToB large-client scenarios. In 2025 we successfully delivered 20+ Agent projects, covering government and top state-owned enterprises, with project sizes from 300,000 to 4 to 5 million RMB, each receiving high recognition. Our approach is relatively "heavy" — an FDE deep implementation model with two teams: an Echo team embedded in client business, translating real needs into AI-implementable solutions; and a dedicated development team building and deploying on our proprietary platform. Essentially, we bridge the gap between foundational large-model capability and real enterprise business, using Agents to help clients "rebuild" their existing operations. In the informatization era, all business had to go on ERP, CRM and other "systems of record," standardizing human operations into databases where the system depended on people. Now with Agents, it's the combination of "person + business system," and operations can run self-service. We never say "replace people" because the true value of Agents is efficiency and revenue growth — it's an engine that helps people run faster, not a knife to cut people.
Guo Yu: We are a platform serving business owners, top business leaders, and ordinary entrepreneurs. Their problem isn't lack of ideas — the problem is that to build software, you need to know design, product, and technology, and manage a relatively large team. Now vibe coding lets them build a product that can go live, operate, use, and iterate through natural language and chat alone. We've delivered to thousands of entrepreneurs who can directly build their own product systems.
Jiang Yaokai: Hello everyone, I'm Jiang Yaokai from Hanzi Tech. We are a no-code development tool serving what's now called the one-person company — actually we started six years ago, when we called it non-technical founders. They need to build a digital solution for their own business, to deliver. What we do is help them build an entire system: from database to workflow to API integration, to Agent construction, to front-end deployment. The goal is to let non-technical founders build their own software and keep iterating it.
Ma Liang: Hello everyone, I'm Ma Liang, founder of Silicon Geek. We mainly do ToC — a personal Agent. Compared to things like OpenClaw and Codex, this Agent is different: it's privacy-first. Why? If users use traditional Agents, they need to provide context themselves, explaining everything from scratch each time. We think that barrier is too high. We actually collect context continuously while the user uses their computer — what software they've used, what operations they've done, what web pages they've viewed — all stored locally. When you give a task, everything you've done on your computer, including who you chatted with on IM, seamlessly becomes your context. In enterprises, when truly landing ToB, local inference also has significant application scenarios.
We defined a term called indirect behavior — the implicit context people produce through daily operations. This context is very valuable because all our daily behavior, including who we contact externally and what tasks we do, doesn't need to be explained to the engine. If you explain to the engine each time, it only has scattered memories from previous conversations — those are just things you previously explained. But more things are generated through interactions with people. We don't want users to have to organize and provide context to the Agent each time; we find that counterintuitive.
Currently it's mainly ToC, but for ToB, overseas it's mainly ToC while we're exploring ToB domestically.
Wu Di: Hello everyone, I'm Wu Di, CMO of Lingzhu AI. Lingzhu is a zero-barrier, zero-code interactive content platform — the first AI product in China positioned as an interactive content platform. Lingzhu's target users are not the 9.4 million programmers but 1 billion ordinary Chinese people. Many people are familiar with the vibe coding sector, but the one that truly lowers the barrier to zero should be Lingzhu. We just started public beta on June 25. Our testers include doctors, lawyers, children, people from all industries, and small business owners. The user range is extremely broad: 60% are aged 30–50, and the second-largest group is children aged 8–12. Our product truly transforms AI from visible code to invisible code. We have huge market space and opportunity and may redefine content platforms. Lingzhu's mission is to be the Douyin of the AI era.
How Do You Judge Product Success or Failure?
Sang Zhuohao: Because the first three companies are more ToB and the last two more ToC. I want to ask: now whether ToB or ToC, project cycles have shortened — ToC projects can be done in 15 minutes. What do your users and clients evaluate you on most? It certainly isn't time or delivery. How do they judge whether your product is successful or failed?
Chai Yatuan: ToB is harder. Many large clients, especially early on, we're constantly filling gaps for large models. Client expectations for AI are very high, requiring perfection, but model capability itself has issues, so much of what Agents do is fill gaps for the model. For client requirements, whether an Agent project can be accepted — whether it meets expectations or the minimum standard — must exceed humans. What humans can do, it must achieve, even better and faster than humans. If human completeness or accuracy is 85%, you must exceed that. Efficiency must be faster: if a person handles 100 orders a day, you must handle at least 1,000 or 2,000. Currently, in ToB, whether we can deliver is the most basic standard for AI projects. At current model capability, this is achievable. If you go back two years — we started four years ago — it was almost impossible.
We first align goals with executives; this is critical. Some previous projects were painful because we didn't figure out what to build at the start. Many ToB clients don't know what an Agent can ultimately become or what they want. We position ourselves as companions exploring together. An Echo team essentially acts as a product manager exploring with the client, understanding the business, and drawing a blueprint based on AI knowledge. First, we sort requirements with business staff, then align goals with CXO-level executives. If that checks out, we continue. During actual delivery, many difficulties emerge, which we ultimately solve together. Despite major challenges, we implement through various engineering means. Engineering implementation is often just patching for the model because models hallucinate and are unstable. For clients, regardless of model issues, the final result must meet expectations.
After extensive exploration, we found that AI can't be required like traditional software. Many clients still use traditional software requirements on Agents — requiring 100% accuracy, predictable results, A in A out with nothing else. This is wrong; if you use AI like traditional software, there's no point using AI. The normal approach is to first understand what AI is — it has creativity. AI output shouldn't be treated as deterministic; it should be a range. As long as you tell AI the limits — through prompts telling it what not to touch, what taboos to avoid — within that range is where AI发挥, which is exactly why we want AI. AI has creativity; it's not mechanical, it has great creative potential. Following this approach, you often get unexpected results. Good AI projects and bad AI projects — currently, if you build a large model like traditional software development, the project will definitely fail, facing massive hallucinations and instability. If you treat it as an AI project and fully leverage AI's creativity, the results are excellent. You must define boundaries clearly upfront.
Guo Yu: The problem we solve is special. Before, we delivered software for enterprises; now we want entrepreneurs to build their own. Ordinary people haven't studied design, haven't built internet products, don't understand technology, haven't experienced product building. How do we help these friends build products themselves? There used to be three ways: hire external outsourcing (very expensive; traditional companies can't afford it); wait for an internal IT team (unimportant business never gets done); or buy SaaS (also expensive, standardized but not personalized enough, every enterprise has its own personalized needs). Now vibe coding lets them do it themselves.
How do ordinary people do this? There are countless problems to solve. First, can they express needs — can AI understand vague, real business needs? Second, can vague needs be designed into a reasonable product? Third, they've never written code, will hit various problems — can these be seamlessly solved? There are countless details and many pitfalls. Making a product is easy; making a product with taste, usability, and stability is hard. Countless things need solving: security and compliance, user permissions, testing, deployment, monitoring, continuous iteration. They haven't experienced any of this before; we must make it so they don't need programming language to get it done. This is what we keep doing, iterating continuously. Ordinary people really just need to chat, and an enterprise can build its own usable, iterable software. The final delivery boundary is when the entrepreneur feels it's truly usable, no extra programming knowledge needed — just understand the business and it's done. It can go live, be used, and continuously iterate, no need to find other programmers.
Jiang Yaokai: We position ourselves as the technology foundation for one-person companies. For a technology foundation to truly work, there's only one condition: you must be able to make money on it — build your business system on top and make money. That's our standard. For example, before AI we already existed, nearly seven years. Around 2021 or 2022, a fund manager who loved collecting sports cards built the first version of a card collector on our platform to track how many cards he had. Other collectors around him started saying they wanted it too, because everyone wants to know what cards they have. As more people gathered, consignment trading, group buying, logistics automation, and grading emerged. He still operates alone; he told me last year's revenue was 6 million RMB. To me, this is the platform's greatest success. As a tool, success comes from users' success — if users make money with it, the tool is successful. This standard hasn't changed with AI; it's just that reaching it used to require someone with strong exploratory spirit and very strong learning ability. He's not technical — he's a PE fund manager — but has very strong learning ability. With AI, the next change needs a new term: not vibe coding but vibe no coding — using AI to help people build on a no-code platform, with lower difficulty, faster speed, while retaining the visibility and understandability that no-code platforms provide.
Ma Liang: For consumer products, especially Agents or all AI products, the golden metric for customer satisfaction is retention. For enterprises, ToB orders and project cycles are long, possibly settled annually. For ToC, it's mostly monthly — how many previously paid or free users still use your product every day next month is undoubtedly the core metric. But this is an outcome metric; you can't just stare at outcome metrics during optimization. You need details: analyze daily usage records, which tasks failed, how to optimize failed tasks, including traditional metrics like activity and payment conversion — all monitored daily. The core, including what investors look at most for consumer products, is retention and process.
Wu Di: Lingzhu's product is special, maybe different from the previous speakers. Our users are interesting — there are doctors, professionals who completely don't understand code and have no technical knowledge. We've always felt that in the AI era, coding brought many programmers in, but locked many more people out — the learning barrier is still very high. Lingzhu wants to bring everyone beyond the 9.4 million programmers into the AI world. Customers are diverse: a sophomore student used Lingzhu to develop a card game, validated it online, got good results, created buzz among classmates, then made the card game physical — orders now approach 500,000 RMB. This completely exceeded product design expectations. Coding isn't just a professional activity for programmers to write software and transform workflows. Chinese people have enormous imagination and creativity; they just lacked tools and platforms to release it. We want to be that bridge.
User feedback is very positive. A doctor at a top-tier hospital used Lingzhu to make a bladder health guide, continuously sharing interactive medical science knowledge to WeChat groups, Moments, and past patients. A 10-year-old's case was featured in People's Daily — in our competition, they made a Pac-Man game, created their own editor and image-editing map editor features, and transformed Pac-Man into a new game. A parent used Lingzhu to create an olympiad math problem collection for their child, turning a mistake notebook into an interactive mini-program to help the child repeatedly correct errors. What we see is that when doing products at big companies, product managers cut requirements to standardize products and make them SaaS-friendly for billing and efficiency. But we see that more people's long-tail needs were cut off and couldn't be realized. Lingzhu wants to be that bridge, releasing everyone's creativity and imagination.
Coding, especially vibe coding — one sentence generating code, truly zero code — only reached the market last year. The market has just begun; the space is enormous and the ceiling is very high.
How Are You Different from Codex and OpenClaw?
Sang Zhuohao: Next, I'm curious. Everyone here is basically building on this generation of vibe coding technology. Jiang Zong has some differences. If you take OpenClaw, Codex, or earlier Cursor as product benchmarks, you must have done something different. I especially want to know what that different thing is and the insight behind it. Why isn't today's product form an OpenClaw or Codex?
Chai Yatuan: Because we have services involved, we also have Agent products and development platforms for different needs. But ours isn't Codex. First, Codex and OpenClaw still require a certain threshold — you need coding fundamentals, and how to compile and use code afterward still has a barrier for ordinary people. They target engineers with professional development ability. Our work may involve some implementation, but mainly serves enterprise business staff or IT departments — we're not in the same category. Like the other speakers, we're pushing no coding, letting ordinary people who understand business but can't write code build things. There's a statistic that only 2% of people worldwide can write code; 98% can't — including Wu Zong from Lingzhu, who probably wants to serve more of those 98%. Codex is evolving too; recently Codex integrated into GPT and is trying to become chat-like. But even so, after writing code, the barrier for ordinary people is still high. You've added a layer of simplification.
Guo Yu: Our clients don't know what OpenClaw or Codex is; they're running their own business and need to build management systems for their operations. These products are all overseas, for programmers, without localized business, data stored abroad, databases built abroad — very slow, and you basically need a VPN to use them. Most products are for programmers. What problems do we solve? A few points: first, through chat alone — even an hour of chatting — AI directly understands their business, asks a few follow-up questions, and after answers, the entire product requirements document is co-created; this process is very smooth. Second, once requirements are confirmed, we provide 4 designs — like home renovation where the contractor presents four styles: Chinese, French, American, classical; if you don't like it, we switch; you find the taste you want and proceed. Design is done by AI after understanding requirements; the user just chooses — they don't design. Third, during the entire R&D process, countless problems arise; we provide a free consultation service called Su Xiaoqiang, an AI assistant you can ask anything — it helps plan features, plan implementation, discuss until satisfied, then immediately executes, guiding you step by step. Programmers do all this — feature planning, understanding phases one/two/three, dependencies, keeping each phase within context — the user doesn't need to know any of it, but we handle it. The end result: a product successfully built and published, all problems solved. We're building a localized Chinese experience so every entrepreneur who understands the business can build great software.
Jiang Yaokai: We don't consider ourselves competitors to OpenClaw or Cursor. We even just released plugins for OpenClaw and Cursor — all three Cs can install our plugin and directly operate our no-code backend. The biggest difference is that they do code generation, which means verification requires understanding code. No-code's characteristic is that logic is presented visually, so even without coding knowledge, you can understand what the system is doing. Software development can be simplified into generating ideas, practicing, and finally testing results. Testing the first two steps — step one is done by the human, whether a fund manager or doctor; their ideas, and AI acts as their hand. Hanzi used to give you a hand you could use yourself and test. Now with AI, there's no need to insist that humans do the operating. But after AI does the final step, if the output isn't on an understandable abstraction layer — if you're a coder you can understand code and dependencies, see the code and know — but if it's presented as code without professional training, it's a black box. A black box is very dangerous for a product owner, because after a certain point, you never know what assumptions exist in the system; those assumptions have sunk into long-ago compressed context. This is the biggest difference from OpenClaw or most vibe coding tools: in the verification step, humans can still participate, and the prerequisite isn't four years of computer science education.
Ma Liang: Setting aside our company's main Agent product, two weeks ago I vibe coded a project myself and open-sourced it on GitHub called Open Vibe Coding — search for it. It's similar to products like Lovable or V0, a vibe coding tool for the mass market. Codex or OpenClaw target technical users more; ordinary users want to build an app but need to think about more things, like how to go live. Products like Jiang Zong's mainly help ordinary users with that — very important: how to launch, handle concurrency, solve security issues, and so on. In the future there's also payments — a mature product and a demo: AI makes demos, 90% of output is demo-level. To reach the final 10%, there's enormous work. This is my personal side project — vibe coded for over a week, built an Open Lovable-like thing and open-sourced it.
Specifically on our own product, it's essentially the same track as OpenClaw and Codex. But our local Agent focuses on ordinary people's office productivity. The biggest difference is that if you have an AI PC — like GB10 or a higher-configured, larger-memory Apple — it unlocks a completely different Agent. Our Agent automatically configures and deploys a large model locally — a smaller model your own computer can handle. It continuously collects operations during daily use, records them locally, absolutely secure. When the Agent actually runs, large model inference is done locally, no cloud model needed, no token anxiety. It continuously reads the screen, captures current browsing content, and uses the local model to distill, summarize, and record. When you need the Agent to work, it knows what you've done before; no need to explain any context. This is probably the biggest difference.
Wu Di: Let me explain. For us, Codex, Cursor and similar tools are for professional programmers doing programs, software, and enterprise efficiency. Our understanding of what coding can create is completely different — we understand that zero code can make interactive education products, interactive game products, interactive literature products, interactive knowledge popularization, interactive psychological assessments. This is unimaginably imaginative. Many people say Chinese people lack creativity and imagination — that's completely wrong. In the two months of product testing, we saw massive amounts of different content products. Our understanding is: coding produces interactive content, and all content forms will change in the AI era. Before, Qidian Novel was for writers; ordinary people couldn't write. But after Qidian, ordinary people can write novels and become bestsellers earning millions a year. In the WeChat Official Account era, WeChat turned many ordinary people into self-media creators. Before, media seemed untouchable; ordinary people couldn't be journalists or editors. After WeChat Official Accounts, technology changed that. With Douyin, Youku and Tudou had been around for years; people thought video was professional, only TV people could do it. But after short video, ordinary people became video bloggers, everyone can create their own short videos, just pick up a phone. In the AI era, content forms will undergo another round of transformation; all content becomes interactive, everyone can participate, create interactive content, produce content. Many people may not produce content but can watch and play on Lingzhu. Beyond zero-barrier AI creation tools, Lingzhu has a marketplace. All interactive content, mini-games, small works, novels, knowledge popularization, and quizzes produced by users are in the Lingzhu marketplace. People who can't create can first play and watch others' works; through modifying and remixing, they can create at a lower threshold — not from 0 to 1, but from 1 to 2.
What Lingzhu does is bring coding's code all the way down to the floor, letting everyone participate and form a marketplace and a growth flywheel. People who can't use it can watch and play, giving creators emotional or material value. Creators can share their knowledge and experience through interactive applications, creating new commercial value. This is how our know-how is completely different from other zero-code companies. It's not just build it for yourself or internal enterprise use; there's also viral spread. Not just spread — it's a reciprocal model. The AI era's zero code has democratized creation rights to the maximum, just as writing rights and video creation rights were democratized. In every era, some right is democratized; in the AI era, it's the right to code. Programmers were always few; many people with brilliant ideas just lacked a programmer. Many had ideas but needed to spend heavily on a technical co-founder. Without one, they couldn't start. Not big ventures — small ventures, OPCs. China has 200–300 million OPC companies and flexible workers who can't code, can't understand what a plugin is, what a ladder or VPN is, what Codex or Cursor is — even the English letters are a huge knowledge gap. Lingzhu bridges this gap; there's enormous potential here.
Which Technology Trend Will Profoundly Change the Industry?
Sang Zhuohao: Last personal question, because I'm also pushing this internally. I want to know which technology trend over the past few months, and the visible months ahead, will profoundly change your product, even this industry? We throw around terms like harness looping, but which do you think has real potential?
Chai Yatuan: Honestly, this is a big question. But yesterday I saw Google DeepMind's CEO Hassabis say AGI should arrive within years. Maybe we don't know if it's credible, but I believe a big trend is coming — models will get smarter and smarter. Whether doing harness, loop engines, or honest engines, it's all patching for current model imperfections. If model capabilities strengthen, the middle work may become easier or less. Whatever comes after loop engine — call it Agent engine for now. Maybe when model capability reaches a certain level, at AGI, as long as you do things well and engineer Agents properly, you can meet all needs.
Guo Yu: Over recent months, the target user group has a characteristic: they don't believe they can build software because they've never done it. But model development over recent months — in understanding business needs, code generation, and sustained output — is getting more stable, so more people believe it's achievable. Thousands of enterprises have already built products with no problems. Models can increasingly produce software that can go live, operate, and continuously iterate — no problem. Three things to do: first, lower the barrier so low that any business-savvy person can build these things. An HR head can build HR software; a finance head builds finance software; a sales person builds sales software — they can definitely do it. Lower the barrier. Second, not just build software but let AI continue participating in operations, growth, and continuous iteration. When they build software, they need to make money, keep operating and iterating — every enterprise need can continue as long as you understand the business. Third, ride the technology wave to make digital capabilities, data capabilities, stability, security, reliability, and complex systems all very smooth — seamlessly get a great system and start using it.
Jiang Yaokai: Speaking of technology, as a coder I need to kick things off. Our product had an Agent builder early on, an assistant called "Koupai" built with Agents. The first wave used LangGraph — a fixed graph where control flow moves between nodes based on large model judgment. In May this year, we spent a lot of time reading leaked OpenClaw source code, and thought about the harness engineering and loop engineering they were talking about — although the names sound very unhuman, there was some truth to them. The biggest change was replacing all of that with a loop-based Agent loop feature. It's not released yet but should launch this month. The focus is letting the Agent make more decisions, giving it context at the right time — at its points of attention, namely the very beginning and end of the system prompt — providing the feedback the system currently needs.
What does "feedback the system currently needs" mean? When building software, the Agent likewise needs to know what's wrong right now. For example, adding a node — does it have missing inputs or outputs? When modifying, can it find where this database table appears elsewhere in the system? When building the front end, can it immediately see the rendered result? Continuously providing feedback in similar ways. Whether human or AI, learning is fastest with timely, rich, non-random feedback. This makes the entire Agent loop performance improve much faster. Adjustment is no longer designing the shape of a graph but continuously optimizing what feedback, what tools, and what instructions each subsystem within the system provides to the AI. This will make product progress much faster over the next six months.
Ma Liang: I'm also a coder; let me share my understanding. Within the next year or two, the biggest perceptible iteration or evolution comes from two things: the base model and Harness stepping on each other's feet, spiraling upward. Actually, neither OpenClaw nor Codex was that good at first; base model capability wasn't enough. But why wasn't it enough? Because they hadn't gotten enough real user cases, scenarios, how users use it, what might fail — that data wasn't sufficient. But as more users find Agents genuinely useful and adopt them, model companies get this data and iterate next-generation models; progress is very fast. Including these past couple of days — Groq 4.5 really has improved too much; after Elon acquired Cursor it's completely different, this new generation of models. This spiraling process is still accelerating, including what people say about AGI — including Zhipu's Tang Jie saying next year, Zhipu might achieve something, so confident. Essentially, the industry finds this spiral accelerating, and the future is worth looking forward to.
Wu Di: From a non-coder's perspective, all technical terms are Silicon Valley technical elites continuously creating new concepts to lead market trends and customer demand. From a zero-code product perspective, what we really care about is how users think, how they experience, how they find it useful. From a product perspective, the product architecture is multi-agent and uses harness engineering. Which technology you use isn't that critical; whether the technology architecture and product design experience truly match the user, and whether users can actually use it, is what matters. Many coding products still require selecting a model — for our user positioning, selecting a model is a huge barrier: is it 4.5 or 3.7, Qwen or MiniMax or DeepSeek? All of this is internalized. We bring users from visible code to invisible code, using a questioning mechanism so users can truly complete creation and generation through dialogue alone. The entire process is internalized in the product — through questioning, continuously drawing and absorbing user knowledge and questions, forming skills, continuously consolidating. The whole process is friendly to ordinary people: no code to see, no skills to understand, no model to know — the backend has already adapted these models. What kind of need, front-end or back-end, the model is already adapted. One sentence from the user, Lingzhu runs itself ragged. The simpler the user experience, the more complex the product and technology architecture behind it.