Luckin's Coffee Agent Skill Shows Why the Super App Era Ends With Tokens

Luckin has also released a skill, which sparked some thoughts about the AI era and the internet era.

The skill is called My Coffee Skill, a zip package. Load it into Claude Code, Codex, or Cursor, and you just tell the agent, "Get me a hot Americano for pickup at the office," and it places the order for you.

Services are decoupling from platforms

In the past, you couldn't order coffee without going through a platform. Whether it was Meituan or some super app, it was the center where you went to order. Platforms were valuable because they aggregated supply and held the traffic gateway.

But when Luckin, KFC, and Mixue turn their own services into skills, every agent can connect directly, and that center gets bypassed. You no longer "open an app to order"; you "let your agent order." The rope between the service and the platform has loosened.

And Luckin doesn't have to pick sides at all. What it wrote is one skill—a SKILL.md file plus a few scripts—that is inherently agent-agnostic. The same package can be read by Claude Code, Codex, Cursor, Qoder, and Kimi. Supporting one more agent costs almost nothing extra.

This is the key point: the reason a service can decouple from platforms so lightly is that a skill is inherently "write once, run everywhere." Before, you had to build a separate integration for each platform, which was expensive and binding; now, one capability works across all agents. The platform is no longer a mandatory path, because bypassing it has almost zero cost.

Why can't platforms monopolize the entry point?

You might think: if services detach from platforms, won't the agents themselves become the new monopoly entry point? Whoever herds everyone into their own agent first becomes the next giant, right?

That's inertial thinking from the internet era. It can't get past one hurdle—tokens.

Looking back at those twenty years of the internet, fighting over entry points and traffic was an iron law. The logic was simple: serving one more user cost almost nothing. Zero marginal cost meant the more users you had, the thinner you could spread your costs; scale itself was the moat. So everyone fought tooth and nail for daily active users and entry points. Bigger meant winning.

In the agent era, this iron law is inverted. Every conversation burns tokens, and tokens are money. If your traffic doubles, the money you burn doubles. Traffic is no longer an asset you can infinitely dilute; it is itself a cost.

Suddenly, "building a big entry point" shifts from printing money to burning money. The more people you corral in, the worse your losses get. So "monopolizing the entry point" simply cannot stand in the agent era: it's not that no one wants to monopolize it, it's that getting big itself will drag you under.

The era of fighting over internet entry points wasn't ended by any company; it was sentenced to death directly by the underlying logic of tokens.

So the entry point will not centralize into a super app again. It will shatter and sink downward—settling onto each person's own individual agent.

Apps don't evolve, but agents do

At this point, someone will argue: what's the essential difference from using an app now? Isn't it still login, payment, pickup—not one step fewer?

The process is indeed the same. But staring at the process misses the point. The point isn't that the interface changed from "tapping" to "conversation"—it's that an app won't evolve for you, but an agent will.

The Luckin app on your phone is the same one on everyone else's phone. You use it for a year, and it still looks the same, doesn't know you, and won't change because of you. An agent is the opposite: the more you use it, the better it fits you. It remembers your taste, proactively asks if you want a cup when the time comes, and can even string together skills for coffee, calendar, and ride-hailing. One sentence—"8 a.m. flight tomorrow"—and it breaks it down into booking a car, ordering coffee, and airport pickup. It's not a tool you use and put down; it's something that continuously reshapes itself around you.

To put it plainly: an app is made for "everyone" and is fixed upon release; an agent is made for "you alone" and keeps changing according to you. That's the watershed.

But it won't magically understand you out of thin air. Behind this is a slower, deeper effort—trust, and time.

Slowly building trust with your agent

No matter how smart an agent is, it still needs a process of building trust with you. To let it spend your money without asking you to confirm every time, to use your identity to place orders across dozens of brands, you have to hand over your wallet. This isn't a technical problem; it's a trust problem. And trust is cultivated, not given all at once.

No one hands over the keys on day one. You let it do small things first, like ordering a coffee. After it gets it right dozens of times, you dare to entrust it with bigger things. Along the way, you casually solidify a rule like "I only drink hot, pickup at the office," and its permission boundaries gradually expand. It's no different from training a new assistant.

The longer you use it, the more rules accumulate between you, and the smarter it becomes for you. It's not that the model got stronger; it's that the "manual of you" you've built together has grown thicker.

And this set of rules, this memory you've painstakingly accumulated, is tied to one specific agent. Switch to another, and you start from scratch. This loops back to the platform question at the beginning: the new moat is no longer how much supply a platform aggregates, but whose house your rules live in. Whoever stores your manual locks you in. The platform hasn't disappeared; it just shifted from "the shared entry point for everyone" to "your private one."

A cup of coffee becomes a skill. Luckin probably just wanted to sell a few more cups, but it incidentally let me see the direction of the next few years clearly: services decoupling from platforms and attaching to agents; the underlying logic of tokens blocking any new monopoly; the endpoint not being a more powerful super app, but an agent you've raised by hand, trust enough, and that connects you to the entire world.

Jarvis isn't built. It's fed, day by day, by you.

---

• My Coffee Skill (Luckin AI Open Platform): https://open.lkcoffee.com/skill