The Harness is the operating system; the model is just a swappable processor
Model companies are refineries that produce base oil — the real money is in the chemical plants and the car companies. Users are more loyal to the Harness: they'll swap the model, but they don't want to swap the shell that holds their memory.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Without Claude Code, Anthropic is a different company
Yusen argues that when Anthropic was feared to be "a company that sells APIs," the real turning point was Claude Code. It brought in a large volume of high-quality usage data, and that data flowing back made the coding model better, forming a data flywheel. Codex later grew fast too, which shows Claude Code and Codex are essentially both Harnesses — it isn't that a strong model lets you sell APIs; users want a product, not an API.
— Dai YusenAnthropic's revenue isn't the return, it's its customers' investment
Yusen breaks the chain into three steps: input, output, outcome. Buying tokens is the input, building software is the output, and the software selling for money or cutting costs is the outcome. Anthropic's revenue surge made many people think the return problem was solved, but that just pushes the problem down to its customers. He asks in return: if a company suddenly had ten times as many engineers, would its revenue surge? So far no big improvement has been observed on the outcome end, because what's missing isn't programmers — it's not knowing what to do.
— Dai YusenThe shell stopped being a shell long ago: model, harness, context, runtime
Yusen uses F1 as an analogy: the model is the driver, and the Harness is the whole set of maintenance and tire changes that keep it on the track. The layers are the model, context, tools and the agentic loop, and runtime/sandbox. What made OpenClaw stunning is that it gave the model a lot of permissions, plus a heartbeat mechanism that checks every 30 minutes whether anything is left unfinished. These mechanisms are all simple to describe, but they make the model complete long tasks better. More and more value sits in these layers wrapped around the model.
— Dai YusenThe Harness is more like an OS, the model more like a processor
Guangmi says model companies are the new-generation OS; Yusen disagrees: the Harness itself is more like an OS, and the model is like the processor that drives the OS. In the Windows era you could plug in Intel or AMD, whichever was cheaper and better; now users can also plug models into and out of the Harness. People building Skills don't have to deal with how to communicate with the model, just as after Windows appeared developers didn't have to handle hardware, only the API.
— Dai YusenNetwork effects may emerge between agents
AI never had good network effects before, because six months ago there was no essential difference between your Claude Code and mine, so there was no reason to exchange. But now each Harness accumulates its own context and skills, and the same task given to different people's agents produces different results. So "my agent hires Zhang Yaojun's agent" may emerge: you pay 10,000 yuan, 1,000 of it is the token value and 9,000 is the proprietary knowledge the other agent has accumulated. This kind of value network is a bit like e-commerce, but it's still very early.
— Dai YusenYour old moat becomes your weak spot
Mobile internet super-apps locked users inside the app to form a closed loop, but if the user's agent can't reach the information inside a closed app, that closed loop becomes a cage. WeChat isn't open; Feishu is open and even shipped a Feishu CLI, so people have an incentive to move work groups from WeChat to Feishu — chat in WeChat and the agent can't see it, chat in Feishu and the agent can. At the same time big companies have high security and quality requirements, while products like OpenClaw carry no historical baggage and can design more AI-native interactions for frontier users.
— Dai YusenAI is like an alien arriving in the human world, in three steps
Step one is giving humans more and better agents (Claude Code, Codex, and Manus are all doing this); step two is making agents adapt to the human digital world — the human data world was designed for humans, and GUIs, captchas, and credit cards were originally meant to block bots, so now we need to issue cards to AI and let agents register and use services as equals; step three is building a digital world exclusive to agents, where for instance payments between agents are high-frequency, small-amount, and many-to-many, and there may be no concept of a card at all, only CLI or API. Every step is a product opportunity.
— Dai YusenThe next ByteDance may not look like ByteDance
Yusen says many AI applications are "AI's information feed": you open it and it's a feed too, users scroll single-column or double-column through things AI produces, essentially wrapping the new technology in the shell ByteDance is best at, then competing on promotion and distribution. Beating ByteDance inside ByteDance's own rules of the game is extremely hard. OpenClaw is another form: it has no app of its own, not even a home turf, but it lives everywhere. Moving the previous era's success paradigm into the AI era is the same as competing head-on with the champion.
— Dai YusenIn their own words · checked verbatim
Anthropic's revenue surge actually made many people think the return problem had already been solved, but my view is that Anthropic's revenue isn't the final return — it's actually its customers' investment.
Anthorpeg的收入大涨 其实让很多人认为 这个回报问题 已经被解决了 但我的看法是 Anthorpeg的收入 它不是最后的回报 它其实是 它的客户的投入
Dai Yusen23:17
I now think the Harness itself is more like an OS, and the model is really like the processor that drives this OS.
我现在觉得 这个Harness本身更像是OS 模型它其实很像是 驱动这个OS的处理器
Dai Yusen52:25
Users can swap the model, but they don't want to swap the Harness.
那用户是 可以换模型,但他不想换Harness。
Dai Yusen59:20
Then your original moat becomes, possibly, a weak spot.
那原来你的护身核 就变成了你的一个 可能一个软肋
Dai Yusen1:21:26
I think it still has to be genuinely useful, not that because I used AI I should be worth a lot of money.
我就觉得还是要真有用 而不是因为我用了AI 我就应该值很多钱
Dai Yusen1:57:46
You'll find that AI today can't tell an original joke — it can only refreeze a joke a human has already told, but it can't tell an original joke.
你会发现现在AI 它没法讲一个 原创的笑话 它只能把一个 人类已经讲过的笑话 给refreeze 但是它讲不出 一个原创的笑话
Dai Yusen2:12:50
Figures
| Codex cost relative to Claude Code | about 50% cheaper | 22:12 |
| Layoff rate at big Silicon Valley companies | 15% | 28:13 |
| Projects invested in last year | looked at 100, invested in 2 | 1:07:30 |
| Projects invested in this year | 3 invested by May | 1:07:30 |
| Last year's valuation of two world-model companies | about 200 million RMB | 1:12:32 |
| A friend's monthly token burn | $10,000 | 2:04:49 |
Glossary
- Harness / orchestration layer
- The whole set of tools, context, agentic loop, and runtime wrapped around the model, making a strong model run by the rules.
- context engineering
- The layer of work that feeds the model the real-time information and organization-specific information it doesn't know.
- sandbox
- An isolated runtime environment that lets an agent go online, use software, and run tasks.
- out of distribution
- Content or problems beyond the data humans already have, that have never appeared before.
- agency
- The ability to decide for yourself what to do and to initiate action on your own — Yusen thinks this is the ground humans have left.
How to listen
For founders and investors trying to judge the boundary between the AI application layer and the model layer, especially those building agents, Harnesses, and AI applications, and product leads torn between big DAU and high-value tasks.
The face-slapping retrospective from 2:00-13:00 can be skipped; the mechanisms and judgments are all in the back half.