Apple’s Cards app was not a failure. It was a two-year prototype for the kind of software everyone is now trying to build with AI agents, and it got shut down for the exact reasons today’s agents keep stumbling.
That is a contrarian take, I know. The mainstream story treats Cards as a cute footnote. Apple launched it in 2011, let you design custom letterpress cards on your iPhone, printed and mailed them for you, and then quietly killed it in 2013 because the logistics and the economics did not work. Filed under: things Apple tried, shrugged at, moved on from.
But if you spend your days explaining AI agents to people who do not write code, as I do, Cards reads very differently. It reads like a blueprint.
What an agent actually is, in greeting card form
When people ask me what separates an AI agent from a chatbot, I usually say this: a chatbot gives you words back, an agent goes and does the thing. You state an outcome. Something else handles the middle.
Now look at what Cards did. You tapped a few times on a phone. On the other end, real letterpress machines ran, a custom heart stamp went on, postal workers carried the envelope, and days later a physical object landed in your mother’s mailbox. You never touched any of that. You did not choose a printer, compare postage rates, or track a shipment. You expressed an intention, and a chain of real-world actions fired off without you.
That is the agent pattern almost exactly. Simple surface, complicated machinery underneath, and a promise that the user never has to see the machinery.
The part that should worry agent builders
Here is where the story gets useful instead of just charming. Cards did not die because people hated it or because the interface was bad. Reports on its shutdown point to logistics and economics. The software worked. The world it reached into was expensive and messy.
This is the most under-discussed risk in the entire AI agent space, and I bring it up constantly with non-technical readers who are being sold on agents doing their shopping, their scheduling, their travel booking.
An agent is only as good as the systems it touches. Those systems have costs. They have failure modes. They have humans in them.
- Every action costs something real. A chatbot answer is cheap. An agent that orders, books, prints, or ships is spending money in the physical world on your behalf. Someone has to cover that, and the per-action math has to work.
- The last mile is not software. Presses jam. Deliveries get lost. Warehouses run out. No amount of clever prompting fixes a supply chain.
- Polish hides fragility. Cards felt refined, which made the underlying operation look more solid than it was. Today’s agent demos have the same quality. A smooth interface is not evidence of a durable system behind it.
Compare it to the Apple Card that stuck
It is a small irony that the Cards app is gone while Apple Card, the credit product launched in 2019, has kept growing and adding features since. Same company, similar naming, wildly different outcome.
The difference is instructive. One product moved paper through a physical pipeline that Apple had to build and pay for. The other moved data through financial systems that already existed at enormous scale. Software plugging into infrastructure that already works tends to survive. Software that has to invent its own infrastructure tends not to.
If you are trying to guess which AI agents will still be around in five years, that is a genuinely useful filter. Ask what the agent depends on. If it rides on rails that already exist and already handle volume, its odds are decent. If it needs someone to build new physical capacity to make the promise true, ask who is absorbing that cost and for how long.
Why a dead app from 2011 earns your attention
I am not nostalgic about greeting cards. I am interested in the fact that we already ran this experiment, at a company with more resources than most agent startups will ever see, and learned something specific: the hard part was never the interface.
Fifteen years on, the industry is rebuilding that same shape with better technology in the middle. The language model can now understand a vague request, plan the steps, and adapt when something goes wrong, which Cards could not do. That is real progress and it matters.
But the constraint has not moved. An agent still has to reach into a world that charges money and breaks in ways no model anticipates. The teams that take that seriously are building something that lasts. The ones that treat it as a detail to solve later are building a very polished version of a 2011 greeting card app.
🕒 Published: