Recently, I helped many friends around me install OpenClaw (nicknamed "Little Lobster"). The entire process was automated using Claude Code; I barely had to lift a finger.
This reminded me of that Tencent video where a bunch of engineers sat there providing hands-on assistance, installing hundreds of units in one afternoon. Claude Code paired with a strong model takes just over ten minutes to complete one installation, costing about ten-plus yuan—perhaps just a few yuan in the future. Remote installation services on the market charge 300-plus yuan; considering each computer has a different environment and various network issues, manual labor taking half a day indeed can't be cheaper. But AI has compressed this cost to one-tenth.
No Method, Only Volume
After installation, many friends ask: How do you use this thing? How do you actually learn AI?
My answer has always been simple: There's no method—just use it extensively.
I subscribed to the Claude Code Max tier with 20x usage limits, basically maxing out the quota every week; one person's usage is comparable to a small team. Use it enough, and you'll naturally know what it can and cannot do, then start getting used to letting it intervene in more and more tasks.
The shift happens at the cognitive level. When encountering problems, the first reaction becomes: Can I get AI to help me with this?
People who haven't reached this stage, no matter how many articles they read, will still revert to old methods when actually working. They can't resist micromanaging and constantly want to supervise. AI cannot perform under this usage pattern.
Worker and Manager
Many people think this wave of hype is like the previous DeepSeek craze—it will pass. I don't think so.
But I do believe most people shouldn't start with OpenClaw right now; they should start with Claude Code.
OpenClaw's positioning is more like a Manager, responsible for allocation and coordination. Coding agents like Claude Code are Workers, responsible for execution.
Connect a good model to Claude Code, and you can feel what execution capability means—give it a vague, complex task, like helping someone remotely install OpenClaw, and it can work continuously for two to three hours, troubleshooting problems on its own until it's done. Seeing this capability, you'll realize you don't need to write complex prompts anymore, nor break down tasks; give it a general direction and it can deliver.
Only when the Worker reaches this level does the Manager become valuable. It can command multiple Workers simultaneously, each task requiring only a simple instruction, then wait two to three hours to see the results.
The Chain Reaction of an Incompetent Worker
Think in reverse. What if the Worker is incompetent?
Give it a vague task, it gets the execution direction wrong, and breaks things in the process. It goes down a weird branch, spinning in circles and constantly seeking help.
What can the Manager do at this point? It doesn't understand the details itself.
It assigned the task to a Worker that couldn't get things done, and the Worker reports back with a blank stare. Both sides discuss pitifully little information in a vacuum; regarding a complex task, the result is definitely a mess.
This is how many people actually feel using OpenClaw right now.
The Economic Calculation
OpenClaw was originally called ClawdBot, connected to the then-strongest Opus 4.5. Opus 4.5 had end-to-end execution capability; that's why the results were so astonishing after connecting it.
Now there are stronger models. Opus 4.6 came out this January, and a couple of days ago OpenAI released ChatGPT 5.4. After using it, honestly I was very shocked—not just a bit stronger, but much stronger.
But the problem lies in the cost.
Claude Code uses the coding plan, similar to a buffet. I pay $200 per month, but the actual tokens used would cost about $2,000 if calculated separately via API—a tenfold difference. These companies are subsidizing, using this method to get technical people on board first.
OpenClaw cannot use the coding plan. It must use API pay-as-you-go billing. For the same usage volume, that's $2,000 per month.
So almost no one on the market connects the most expensive models to OpenClaw. For convenience, everyone connects domestic models.
Domestic models are indeed improving fast, but there's still a gap compared to the top-tier ones.
If the Worker isn't strong enough, the Manager becomes a decoration. Many people feel OpenClaw is like a toy.
But This Won't Be a Flash in the Pan
Model progress is much faster than imagined. Opus 4.6 came out in January, and two months later ChatGPT 5.4 surpassed it. Domestic providers will catch up quickly too.
The infrastructure is already deployed; large numbers of people have installed OpenClaw. When new models come out, you only need to switch the connection.
Last month it still felt a bit dumb; two months later you switch the model and find it's a completely different species. In half a year, at current speeds, ordinary people will be able to experience that sense of shock from the strongest models.
Moreover, the barrier to entry is indeed low. Install OpenClaw locally, spend a few dozen yuan per month on API fees; domestic cloud vendors offer subsidies, ranging from twenty to forty yuan. When DeepSeek 4 comes out, connect it and you can use it immediately. No longer just a chatbot.
OpenClaw lets people who don't write code use AI agents too. Everyone knows how to use a chat interface; the barrier is minimized.
It may not be the best experience currently. But everyone has already glimpsed through the door crack what AI can do. As models upgrade, this door will open wider and wider. Install it first, start using it first.
Originally published at https://guanjiawei.ai/en/blog/worker-before-manager
Top comments (0)