Andrej Was Right About Ephemeral Apps | Elacity Labs
Karpathy spent an hour vibe-coding a custom cardio dashboard. He says it should have taken a minute. The reason it didn't isn't the LLM — it's the operating system underneath. The App Store is the wrong shape for the agent era. Here are the six primitives the one-minute app actually needs.
Andrej Was Right About Ephemeral Apps: The OS That Has to Exist for Them to Ship
Andrej Karpathy posted a thread that ate the timeline. He had spent an hour vibe-coding a custom cardio dashboard for his own 8-week heart-rate experiment. Claude reverse-engineered the Woodway treadmill's cloud API, pulled the raw data, processed it, and rendered a working UI. It worked. It took an hour. It should have taken a minute.
He is right.
The reason it took an hour is not the LLM. The reason is the operating system underneath. Mobile computing is still organised around a 2008 primitive — the discrete, downloadable, App Store app. The agent era needs a different primitive.
The App Store Is the Wrong Shape
The App Store assumes there is a long, stable list of programs that millions of users will want. It made sense when software was expensive to write. It does not make sense in a world where a model writes a working program from a natural-language sentence in fifteen seconds.
Karpathy's point is sharper than "AI will replace developers." His point is that nine out of ten apps simply will not exist. They will be summoned, used, and thrown away — bespoke software for a bespoke task, written for a single person, for a single afternoon, for a single experiment.
There will not be a Cardio Experiment Tracker in the App Store. There will be no Cardio Experiment Tracker anywhere. There will be a one-line ask — "help me track this for eight weeks" — and a working app a minute later. Then it disappears.
Why It Is Not One Minute Yet
The bottleneck is no longer model latency. It is the substrate the agent runs on.
For the one-minute app to ship, six primitives have to be real and reachable from a single agent prompt:
- A Wealth Capsule that holds your personal context, identity and history — so the agent does not start every session from zero.
- A sensor catalogue that lets the agent discover what is wired to you — treadmill, watch, camera — without reverse-engineering each vendor.
- A dDRM layer that lets every called engine charge or license its work cleanly.
- A Personal Cloud Compute substrate that can host the agent itself, on your own hardware.
- A programmable payment rail — stablecoins on x402 — so the agent can pay sub-cent prices without a card.
- A location-agnostic OS that decides at call time whether a workload runs on your phone, on your home node, or in a cloud enclave.
We have shipped four of the six. The other two — the sensor catalogue and the registry of agent-native services — are what the industry has to build next.
The Sensor Bazaar That Replaces the App Store
Karpathy's second point cuts even deeper. Every device he interacts with — the treadmill, the watch, the heart monitor — is fundamentally a sensor. It turns physical state into digital knowledge. None of them ship with an agent-native API. All of them ship with a human-readable PDF and a portal you log into with a password.
This is the next great industry rewiring. Every appliance, every API, every SaaS product becomes a headless engine with a machine-readable schema. The marketplace that replaces the App Store is not an app marketplace. It is a sensor and actuator bazaar — pieces an LLM can compose on the spot into apps that did not exist five minutes ago.
That is exactly what we mean by the Death of the App. The app is not dying. The shape is dying. The discrete, downloadable, screen-by-screen container is being replaced by a temporary script that wires together a few engines and a sensor or two.
What Elacity Is Building For It
The Capsule is the unit that survives the transition. It is the only artefact in the chain that stays around — your data, your licences, your identity — while the apps that operate on it appear and disappear.
When you ask for a cardio dashboard, the agent does not download an app. It reaches into your Capsule for your heart-rate history, your sleep data, your past sessions. It composes a UI that exists only as long as you are looking at it. The data stays in the Capsule. The session ends. The app goes away.
When the same agent next handles a different task — a tax summary, a holiday booking, a research note — the Capsule is the part that persists.
One Minute, Not One Hour
Karpathy's hour-long vibe-code is the high-water mark of the present. The next year is about collapsing it. The reason this is interesting is not the heart-rate dashboard. It is the implication: that the entire SaaS industry — every paid monthly subscription you have ever signed — is built on the assumption that you would never compose your own version in a minute.
That assumption ends in 2027.
The companies that ship sensors and engines first will own the bazaar. The companies that still ship apps will own a museum.
We are building the substrate that lets the one-minute app happen at all.
Run an Elacity node from your home.