The API is the product. The UI is a skin.

Cristiano Sacchi

Last week I argued that AI agents should come in through the door, not the window: use the API a service built for software, not the screen it built for humans.

This time I want to go one step further. I think the door is going to become the building. Tomorrow's apps will be APIs, and the interface the vendor ships will be just the default skin.

Not every app, though. And the best way to see where the line falls is shopping. Last week's example was Amazon blocking Meta's Muse, an AI agent that shops on your behalf. Amazon's position deserves a closer look, because it's more interesting than "big company blocks agent."

Not everything you buy is a light switch

Amazon's worry is legitimate. An agent that shops for you skips the ads, the recommendations, the "customers also bought"... and compares prices without blinking. That squeezes margins.

But I suspect the worry only holds for part of what Amazon sells.

I recently bought a light switch on Amazon. It was a chore, and not because Amazon made it hard. It was boring. I bought it because I had to, not because it would make my life any better. Shopping for a light switch is about as fun as shopping for insurance. (Insurance does make your life better when something bad happens... but that's another story.) I would have loved an agent to do it for me.

Now: would you let an agent buy you a dress? A chair for your living room? Hell no. A chair may be needed too, but it's also part of your home, and a good one makes your life a little better. A dress may not be needed at all, but it makes your life a little better every time you wear it. And you won't delegate those choices. Not because you don't trust the AI. I wouldn't delegate buying a chair to my best friend. Because I don't trust him? No... because we don't have the same taste.

So there are two kinds of purchases. Commodities: things you need, where one unit is as good as another and the choice comes down to specs and price. And taste, where the choice is you.

Proof that commodities get automated? Stocks. One share of a company is the same as any other share, and stocks have been traded by algorithms for decades. Nobody misses clicking "buy" at the right second.

So if you sell commodities, AI shopping is coming for you, and you'd better give the agents a door before they come through the window. If you sell anything driven by personal taste, keep investing in an interface that shows your products as realistically as possible... and sleep well.

Today it's good engineering. Tomorrow it's the architecture.

Building an app entirely on top of its own API isn't new. It's just good design: the engine does the work, and the UI is one client among others. Amazon itself did this internally around 2002, when Jeff Bezos told every team to expose its data and functions only through service interfaces. Stripe built a business where developers mostly meet the API, not the dashboard.

What AI changes is that this stops being a nice-to-have.

First, agents. An agent that talks to an API (or MCP, which is the same idea with a self-description attached) is fast, scoped and reliable. If every feature of an app is in its API, there's nothing left for a screen-driving agent to do, except legacy apps and the odd corner nobody wired up.

Second, custom UIs. Until recently, a bespoke interface was something only large enterprises could afford. With AI coding, a 10-person firm can build its own screen for the three things it does every day, and add the features the vendor never will. The vendor's UI becomes the default skin: good enough for most, replaceable for anyone who wants something better.

I don't think this is the only architecture going forward. I do think it will be the dominant one for enterprise software and for any app that isn't heavy on interaction.

Separation, not location

The obvious objection: what about apps that are all interaction? CAD, Photoshop, games?

The answer is that API+Skin is about separating the engine from the skin, not about putting the engine in the cloud.

Across the network, it already works for demanding apps. Onshape is a full 3D CAD system: the geometry is computed on cloud servers, and the interactive 3D model lives in your browser. If CAD can be split this way, your CRM certainly can.

On the same machine, the API can be local. A small "ambassador" agent on your computer talks to the app's local API. No network, no latency. That's not science fiction either: Word and Excel have had a local automation API for decades. It's why macros exist.

And games already do it. They just call it modding. World of Warcraft has let players augment or replace its interface with add-ons written in Lua since it launched in 2004, and for many players the default UI is just a starting point.

Now picture a 3D Dungeons & Dragons game built this way. Players use the local API and an AI to design their own characters... describe them, tune them, give them a backstory... and then play them in the game. Modding used to require a hobbyist programmer. With AI, it requires someone who can describe what they want.

The honest limits

Some products are the UI. Photoshop and Figma have plugin APIs, but nobody buys them for the API. There, the default skin is not a trivial thing to replace.

And apps that make money from your attention will resist. In 2023 Twitter ended free access to its API, and Reddit priced its API high enough that popular third-party apps shut down. Roughly: if the user pays, the app has every reason to open up. If the user is the product, it will fight.

Where Heurivon fits

Heurivon is already a skin, but not a thin one. The voice and chat interface is the part you see. Underneath, it runs on other companies' APIs: Gmail, Calendar, QuickBooks, Shopify, HubSpot. Those APIs on their own are just doors. What makes them useful is what sits in between: the AI that coordinates the work across services, and the harness that keeps it predictable, lets deterministic code do the executing and keeps the AI to interpreting what you want.

It only works because those companies built doors. The API gives you the parts; the value is in putting them together.

Good engineers have built software this way for years because it was clean. AI is about to make it the default because it's necessary.

— Cris