The Million-User Gap for Coding Agents
Jensen Huang has been repeating the same message these past two years: "The age of agentic AI is here," "a trillion-dollar opportunity." In October last year, he said every NVIDIA engineer was using Cursor. At this year's GTC, he painted a picture of 75,000 people working alongside 7.5 million agents.
It sounds like coding agents are already everywhere. I went and checked the actual numbers.
The Numbers Are Surprising
Claude.ai has roughly 10–20 million monthly active users. Third-party estimates put Claude Code at about 1.6 million weekly active users. Cursor has over 2 million users, 1 million of them paying. On the open-source side, OpenCode has 140k stars on GitHub, and Cline has been installed more than 5 million times on VS Code.
Those aren't small numbers. But GitHub Copilot alone has nearly 20 million users, and a JetBrains survey from early this year found that 74% of developers are already using some kind of AI coding tool.
Copilot is mainly autocomplete—it doesn't count as an agent. The ones that truly work in agent mode—you give them a task, and they read files, write code, and run tests on their own—add up to maybe a few million to just over ten million users combined.
Something called "world-changing" by the highest-market-cap company on the planet has user numbers at this scale. I would've guessed at least tens of millions.
The picture in China is different, though. The Kimi platform has over 30 million monthly active users, and ByteDance's Trae has more than 6 million registrations. But those numbers include a lot of non-programming usage; the actual number of people using coding agents isn't that high.
Many people are debating how to configure Skills or wire up MCP. But maybe we need to take a step back: why can't so many people even take the first step?
It's Not Just About Writing Code
The most common misunderstanding is treating a coding agent as a "tool for writing code." People who don't code think it's irrelevant; people who do code think it's just an upgraded Copilot.
In reality, it does far more than code. Splicing videos, batch-processing files, scraping data from websites, debugging unfamiliar software—it can handle all of that. In plain terms, it helps you control your computer to get tasks done; code is simply its operating language. I wrote before that this is a meta-ability—a way for ordinary people to truly take control of their own computers with a very low barrier to entry.
Mobile phones are the exception. Phones are inherently designed around GUIs, so coding agents don't work well on them. Those earlier projects that used models to control phones got some hype for a while, but the approach is completely different and still not quite there.
The NPC Woke Up
The deeper issue is mindset.
When it comes to using software, people have grown accustomed to pre-designed interactions. Click here for this, drag there for that—everything has been designed for you. NPCs in games work the same way: they give you three options to choose from, and whether you read the dialogue or not doesn't matter.
Now the NPC can suddenly think. It waits for you to speak, then acts accordingly.
This feeling is exactly like when ChatGPT first came out. I made quite a few tutorials teaching people how to use it, and later realized most people got stuck on expression. They felt they had to become "prompt engineers"—speaking with precision, making the AI obey, and maybe throwing in some fancy tricks for advanced use.
It's not that complicated. Treat the coding agent like a colleague. It understands what you say and can look through the files in your project. If you clearly explain what you want to do, you're mostly done.
Don't Do It Yourself
There's another trap tied to habit.
The more capable you are, the easier it is to fall into it. When you run into a problem, your first instinct is to fix it yourself. It's like being a manager—even though someone else could do it, you always feel it's faster to do it yourself.
But a coding agent might be ten times faster than you. The quality may lag for now, but the iteration speed makes up for it. The problem is, once you start doing it yourself, you slide back into the old routine—ask DeepSeek or Doubao for advice, tweak it yourself, and use the AI as a consultant. You're still the one doing the work.
These days, I hand 95% of my computer work over to a coding agent. Research, writing, programming, sending emails, interacting with web pages. Once you cross that threshold, you don't need anyone to teach you, because you can have it figure out how to use itself better. That's the meta-ability. I built this website from scratch exactly this way—no front-end knowledge, no DNS knowledge, no SEO knowledge; I let Claude Code handle all of it.
Human in the Loop
But don't be too optimistic either.
Concepts like agent managers and multi-agent orchestration sound compelling—let AI manage AI, running fully automated. The direction is right, but we're not there yet.
In practice, having a knowledgeable person in the loop makes a huge difference in efficiency. If you let agents orchestrate themselves completely, things go wrong once the task gets complex. I talked about the relationship between Worker and Manager before—this is exactly what I meant.
For the near term, someone still needs to be in the middle. But that person's role isn't to do the work with their own hands; it's to clearly state what they want, glance at whether the result is right, and make the call at key moments.
Cross this step, and many things naturally fall into place. Fail to cross it, and you'll stay on the outside watching other people use it.
Originally published at https://guanjiawei.ai/en/blog/coding-agent-adoption-gap