UI SDK
Justy and Cody dig into Vercel's UI SDK and land on the real move: giving agents reusable interface behaviors instead of finished components. They like it less as a component library story and more as a legibility story for codegen and UI automation, while noting the article's speed and cost claims are intriguing but thinly evidenced on the page itself.
Transcript
Justy Okay, this one's Vercel, and the pitch is that the agent should NOT start from a finished Select. It should start from the guts of a Select.
Cody The guts being the interesting part, honestly. They're turning UI interactions into behaviors. Popup, selection, typeahead, roving focus, reorder, resize.
Justy Which sounds weird for about ten seconds.
Cody And then you realize that IS the product claim. If the agent learns those pieces, it can build the obvious control and the next thing you haven't named yet.
Justy So a behavior, in their language, is just the interaction logic. The moving parts under the painted button, not the button.
Cody Arrow-key navigation, focus management, screen reader support, drag and drop. That layer.
Justy And honestly this hit my favorite button immediately, because it's the reusable layer where UI usually gets fiddly and expensive, instead of another look-our-agent-made-a-pretty-demo thing.
Cody Your little selector-on-a-toaster heart is thrilled. But you have a point, it's mechanistically cleaner than shipping one hundred magic components and hoping the model copies them correctly.
Justy Their Select example is basically the whole case. Give the model createPopup, createSelection, createTypeahead, and suddenly Select is one composition.
Cody But so is autocomplete. Combobox, token input, whatever awkward hybrid your app actually needs. And they say they've built more than fifty of these behaviors.
Justy Fifty? Already?
Cody More than fifty. And the examples on the page run broad. Roving focus, virtualizer, resize handles, reorder, and the, uh… the pan and zoom one…
Justy Viewport.
Cody Viewport, thank you. Plus the calendar-ish stuff through the starter components.
Justy And then the cross-framework part. They say the same behavior system works in React, Vue, Svelte, Solid, and Web Components. So you don't relearn interaction logic every time your stack changes… that's the half of the pitch beyond generate-this-once.
Cody That part holds up technically. Interaction patterns really are more stable than component APIs.
Justy Mm.
Cody Abstract at the behavior layer and you get a clearer target for the agent, a more inspectable one for humans. Simpler prompt layer, more structured execution layer.
Justy Which matters most for teams letting agents touch real product UI, not just internal CRUD screens. If you've ever had an agent get eighty percent there and then absolutely faceplant on keyboard behavior or dismissal logic, you know
Cody Aimed at THAT pain, yes.
Cody And I like that they don't stop at build. The claim is agents can use the controls directly after they're built, by reading the page structure instead of relying on screenshots.
Justy In plain English, the agent interacts through the underlying semantics instead of trying to visually guess where the button is.
Cody Which is a nice symmetry, actually. Built by behaviors, used through behaviors.
Justy So much more appealing than the current ritual, where one model generates a thing and then a second model squints at pixels trying to operate it.
Cody Very "I assembled the machine and immediately forgot where the switch went."
Cody The one place I actually slow down is their performance line. They say more than twice as fast and less than half the cost… and the page shows no setup, no task set, no baseline.
Justy Same, weirdly. I buy the mechanism before I buy the number. Less bespoke glue, semantic operation, sure, faster and cheaper is plausible… I just can't tell from the page how general that is.
Cody Yeah, yeah.
Cody So separate the claims. Strong abstraction idea, pretty convincing examples, open-source beta repo through Vercel Labs… and the benchmark-y part is a teaser, not evidence, at least on this page.
Justy There's also the adoption thing. Headless UI for humans was already useful. Headless UI for agents might be MORE useful, because agents are so bad at hidden interaction rules. Make the rules legible and codegen has fewer ways to be subtly wrong.
Cody Mm.
Cody And accessibility is lurking inside that. Roving focus, keyboard navigation, screen reader support, popup dismissal… exactly the details people skip when they're vibe-coding a component at two in the morning.
Justy Speak for yourself, some of us vibe-code at very respectable hours. But that's why it feels real to me. Actual UI craft, turned into reusable infrastructure the agent can reason about, instead of another shiny widget pack.
Cody Which, to be fair, is a much better use of agent energy than asking it to memorize every framework's weird little component folklore.
Justy UI SDK is a good name for that, annoyingly. Episode eleven seventy-eight, and Vercel made us say something sensible again. Rude.