The customer is not the human anymore

Karpathy makes the same argument twice in one interview — smart-home apps should be APIs, docs should be markdown for agents rather than HTML for people — because in both cases an agent consumes the interface and routes to a human. Where that lands for this notebook, where it thins out, and what changes when a vendor does it deliberately.

From the NoPriors interview. The line is his: the industry has to reconfigure because the customer is no longer the human, it is agents acting on behalf of humans. He arrives at it twice, from two unrelated directions, which is what makes it worth a page.

First pass: apps should be APIs

Having replaced six smart-home apps with one text thread (see The claw layer: an agent that persists when you close the lid), his conclusion is that those apps should not exist. Expose endpoints; let the agent be the glue. His argument for it is not tidiness but capability: an agent calling several systems can do automations that no individual vendor's app will ever offer, because no vendor's app knows about the other five. His other example is a treadmill whose app he has to log into through a web flow to find out how often he did cardio.

The objection he anticipates is that this requires ordinary people to vibe-code their own tooling. His answer is that today it does, and that this is a statement about the current moment rather than the shape — the barrier drops, and what is left is ephemeral software the claw assembles and presents without you being involved in how.

Second pass: docs should be for agents

Much later, talking about micro-GPT, he gets to the same place from teaching. If you maintain a library, the instinct is to write documentation for the humans who use it. His version: write markdown for the agents instead, because if the agent understands the library, it will explain any part of it to any person, in their language, at their level, three different ways, with unlimited patience. The human-facing explanation is downstream and can be generated on demand.

He is describing this as something already happening to him. He started making a walkthrough video for micro-GPT and abandoned it, on the grounds that 200 readable lines plus an agent that can explain them is better than his video. What he thinks survives is curriculum: a skill file that scripts the order in which the agent should walk someone through the code — start here, then this — because the sequence is a judgment the agent will not make as well.

The sharper version of the same point is what he says about writing micro-GPT itself. He tried to get an agent to produce it — boil a neural network down to its essence — and it could not. It gets the result completely once it exists, and can explain why every line is the way it is, but it did not arrive at it. His conclusion is a rule for where to spend time: the things agents cannot do are your job now; the things they can do, they will shortly do better than you.

What this notebook already agrees with

This is "what the agent can't see doesn't exist" arrived at independently, and it is worth noting that the two statements come from opposite motivations. OpenAI's is defensive — put it in the repo or the agent cannot use it. Karpathy's is a claim about audience — write for the agent because it is the better distribution channel to the human. Same instruction, different reason.

It is also, fairly directly, the bet this notebook runs on: markdown in a git repository, maintained by agents, with the operations themselves written as markdown files in .claude/commands/. See How this notebook works.

Third pass, from the other side of the counter

Four months later, DoorDash's two co-founders arrive at the same conclusion from the vendor's chair, which is the seat missing from Karpathy's version. Their evidence and their moves:

• Ask DoorDash, a conversational surface over the catalogue, where half of restaurant trajectories end at a merchant the customer has never used and grocery baskets run ~40% larger. Numbers with no denominators — see the source note — but they are the first figures in this notebook attached to the claim that a natural-language surface changes what gets bought, not just how it gets found.
• A DoorDash CLI, launched July 2026, described in exactly Karpathy's terms: reduce the friction for an agent to participate. The worked example is an office manager who pointed a camera at the pantry shelf and has the agent restock when it looks empty.
• The stat Andy Fang says he keeps returning to: more agent traffic on the web than human traffic (asserted, unsourced), and the inference that if you started DoorDash today it would be agent-first.

What the vendor case fixes, and what it doesn't

It localises the objection below rather than refuting it. DoorDash's revenue is the transaction; the app is a means to it. A vendor like that loses nothing by being callable and gains order volume, which is why it ships the CLI unprompted. A vendor whose app is the business model — subscription, upsell, ad inventory, data — has the opposite incentive. So Karpathy's argument holds exactly where the interface is not the product, and the treadmill company he complains about is on the wrong side of that line. That is a sharper claim than his, and it predicts which endpoints will appear without anyone being persuaded of anything.

It is not the shape he asked for, though. Ask DoorDash keeps the conversation inside DoorDash, and the CLI is DoorDash's own agent surface: the storefront rebuilt for agents, not dissolved into endpoints that somebody else's claw composes. The capability argument in his first pass — the automation no single vendor's app can offer, because it spans six of them — requires the vendor to accept being one call among several. A marketplace that owns routing, merchants and a fleet has every reason to be the composer rather than the composed. Which way the CLI goes is the thing worth watching, and the interview does not say.

And nobody has discussed the credential. Karpathy withheld email and calendar from his claw on the grounds that it was too rough to hold his digital life. A shelf-restocking loop needs a payment method on file, and it is the vendor shipping the tool that makes it easy. The failure mode of an unattended loop that spends money is absent from both interviews — see The claw layer: an agent that persists when you close the lid, where the same gap shows up as an unaddressed blast radius.

Where it thins out

Somebody has to pay for the API. The apps he wants deleted are, for most vendors, the surface where the business model lives — subscriptions, upsells, data. "It should just be endpoints" is an argument about what is technically better, and it does not engage with why the apps exist. The Sonos anecdote is not a counter-example; it is a device that shipped without authentication. The DoorDash case above is the real partial answer, and it is partial in a specific way: it shows the argument working where the app was never the product.

The claim is made by someone who can build the replacement. He can vibe-code a home-automation dashboard in an evening, so for him the UX layer genuinely is disposable. It is worth being careful about generalising from an operator at the extreme end of skill to what the barrier will look like when it drops — that is the same inference the interview makes about coding generally, and this notebook has already logged a reason to be suspicious of it.

And the docs argument has a dependency it does not name. "Write for the agent, it will explain it to people" is only true while the agent's explanation is faithful. There is no check in the loop: a wrong explanation of your library reads exactly like a right one to someone who came to the docs because they did not know. The version of this in the notebook — hub pages and human-readable notes that agents maintain but people read — keeps the artifact human-legible on purpose, which is a hedge against exactly that.