AI Coding: A Meta-Ability
Yesterday, while introducing our product to a friend, I suddenly realized something.
Underlying technology is evolving rapidly—how rapidly? It's changing by the day. But the general public—even many white-collar professionals—are adopting these technologies far more slowly than imagined.
Take a recent example. OpenClaw became popular, and everyone was captivated by this technology. Someone asked me: "Are you using Claude Code? Is that for helping people write code?"
I said yes.
Their first reaction was almost identical: "This sounds complicated. I probably won't need it in my lifetime."
Not What You Think
I demonstrated in the most intuitive way: installed in 5 minutes, works with just an account, as long as you can speak.
They stared at my terminal for a while and threw out a series of questions. Is installation complicated? Is it difficult? What problems can it solve?
I said, installation is just one line of code, 5 minutes. An account costs just a few dollars a month to get started.
They froze there. "Is that so? I thought something this powerful must have a complicated installation process."
There isn't.
What's even more surreal is OpenClaw itself. Online, there are circulating 15-step, 20-step installation guides that look headache-inducing. Many people complain: the crayfish is quite useful, but installation is too difficult. On Xianyu (a Chinese second-hand marketplace), there even appeared services charging 500 yuan for door-to-door installation.
I demonstrated on the spot. When your computer has Claude Code installed, you only need to say one sentence—"Help me install OpenClaw on this computer."
It will find out what OpenClaw is, where to download it, and how to deploy it on its own. You tell it what your Feishu (a Chinese enterprise collaboration platform) App ID is, and it will help you debug and test until it runs. You go get a cup of coffee, and when you come back, it's deployed.
One line of code, 5 minutes. That simple.
A Meta-Ability
Someone asked me: "Does your software have out-of-the-box applications? Like text-to-image?"
I didn't answer directly. I was thinking about something else.
If we let everyone master an efficient AI coding tool—a coding agent—whatever you want, have it install it for you.
Text-to-image? You say: "I want a text-to-image application, download the model and deploy it locally, so I can generate images with a single sentence." It might spend an hour or two helping you download the model, deploy it, and get it running.
This capability has already transcended the concept of "application."
It is helping every ordinary person—with an extremely low barrier to entry—truly control their computer. By "control," I mean to a degree exceeding that of 90% of engineers and senior engineers. And your computer is connected to the internet.
So the previous notion of "I need to install an out-of-the-box application" is actually unnecessary. You need something, just tell it. It will do it for you.
What Can It Do?
Honestly, now when I use a computer without this kind of software, I don't know how to use the computer.
It can utilize 90% of the capabilities on your computer that you normally don't use.
You have 20 Excel files to process? Put them in a folder, throw it at it, and tell it how you need them analyzed. It will invoke Python, download plugins if they're missing, and solve problems on its own.
Want to implement some interesting network functions? It can help with that too.
As long as the model is strong enough and fast enough, these tasks that originally required professional skills become within reach.
What Does the Future Look Like?
In the future, people will become increasingly lazy when using computers, unwilling to manually operate them. I rarely actually operate the computer now—at most, I open a few Claude Code instances and switch between them to let them handle different tasks.
Work methods will change.
Before it was: Human → Tool → Computer.
In the future it will probably be: Human → Agent → Computer.
What's the point of building a complex interface? It's a burden for AI, and humans don't look at it either. Just give instructions to the agent and let it execute. Done.
Software infrastructure might change. I'm not sure what it will become, but the general direction is probably this direction.
The Difference Between Claude Code and OpenClaw
Many people ask me this question.
Strictly speaking, OpenClaw is a layer built on top of Claude Code's capabilities. Some call this layer the "lobster layer"—I think that makes sense.
What Claude Code does is: you give a vague instruction, it executes end-to-end, basically meeting expectations during the process, without needing you to correct course, can run for tens of minutes or even hours, and finally gives you a result.
What OpenClaw does is: add another layer on top—distributing instructions, managing results, forming long-term memory, like a project manager or supervisor. It can interact with you at a higher level and become your personal assistant.
But the premise of all this is: the underlying coding agent executes beautifully on every task, and doesn't require your supervision.
This is also why OpenClaw connected to different models is a completely different species.
Currently, there aren't many models that can truly achieve "execute well from a vague instruction." I've only seen Claude's Opus, and possibly Codex's latest model, capable of this. Versions connected to other models feel like toys—every execution is uncertain, making the coordination layer above even more difficult.
The reason OpenClaw can amaze everyone is because it's connected to Opus. But Opus's coding plan is now discontinued, too expensive.
This is the current pain point. I don't know when it will be resolved.
Originally published at https://guanjiawei.ai/en/blog/coding-agent-meta-ability