Luckin's Coffee Agent Skill Shows Why the Super App Era Ends With Tokens
Token-based marginal costs invert the internet's core scaling assumption. Every additional user conversation burns money, so the winner-take-all platform model collapses economically. Developers building agent tooling should watch where user-specific rules and memory get stored, because that storage becomes the real switching cost.
Luckin Coffee released My Coffee Skill, a zip file containing a SKILL.md and scripts that lets any compatible coding agent place an order. The same package runs on Claude Code, Codex, Cursor, and others with no per-platform integration work. This write-once-run-anywhere property lets services bypass aggregator platforms entirely, since supporting another agent costs almost nothing.
The deeper shift is economic. Internet-era platforms chased scale because serving one more user had near-zero marginal cost. Agent interactions burn tokens on every turn, so traffic is a direct cost, not an asset to dilute. A monopoly entry point becomes a money furnace, not a moat. Entry points will fracture downward into personal agents, each one accumulating rules and memory specific to its owner. The new lock-in isn't supply aggregation; it's whose agent holds your preferences, permissions, and history.
The argument that token costs kill the super-app model is a direct economic claim, not a metaphor. If inference costs don't drop to near-zero, the math genuinely forbids a free-to-grow entry point.
Skill portability cuts both ways: services escape platform lock-in, but users become locked into whichever agent holds their accumulated preferences and permissions. The battleground shifts from distribution to memory custody.
Luckin's skill is a real, downloadable artifact, not a concept demo. That a major Chinese chain shipped this now suggests the agent-as-interface pattern is moving faster in China's retail sector than in the West's.
The discussion splits between those who see the agent skill as a business-model play to bypass super-app tolls and those who doubt its practical appeal. On one side, decoupling from Meituan cuts traffic-distribution layers; on the other, platform perks still bind merchants, and typing tokens to order coffee feels like friction compared to a few taps. Trust, liability, and cost dominate the objections — who pays when the AI gets it wrong, and why burn your own tokens for something a cheap app already does.
Luckin selling coffee is essentially multi-level traffic distribution and cross-channel referrals. Using skills to decouple from platforms like Meituan reduces the layers of traffic distribution — that's a business model.
Even if Luckin wants to decouple from traffic platforms like Meituan, as long as the platform occasionally offers perks, physical store owners still have to rely on the platform to sell products. The author's point that 'Agents can form user behavior memory' could save costs beyond just money — that's a highlight — but compared to saving money, its support and appeal aren't that solid.
It depends on the client-side presentation. If it's on a mini-program, it wouldn't be restricted by the Meituan platform.
Alright, but how can you trust an AI to directly take over your account? If something goes wrong, how do you get compensation?
I suggest you try using AI first. The processes are all designed by people, so if problems arise, people are responsible. Don't shift the blame onto AI.
Exactly. If a dish cooked in a wok gets burnt, you can't blame the wok.
It won't work. I can just tap with my fingers to order a coffee — why would I type and burn tokens? An agent app with no irreplaceable value.
In the future, smart butlers — rich folks won't mind spending a bit more for that.
Rich folks aren't idiots. Just like that food-ordering thing from Qwen, I uninstalled it right after using it.