I spent about an hour of WWDC watching Apple tell developers to make their apps more beautiful.
Liquid Glass everywhere. A new transparency slider. Refined edges, better legibility. And this year you don't get to opt out. Recompile with Xcode 27 and your app adopts the new look whether you asked for it or not.
Then, in more or less the same breath, Apple told us the look might not matter much longer.
Because the other half of that keynote was App Intents. New entity and intent schemas that let your app push its content into Spotlight's semantic index, so Siri can find it and act on it through plain language. A View Annotations API so Siri can reach into what's on screen and do something with it. Foundation Models going open source. Xcode itself running coding agents from Anthropic, Google and OpenAI, wired in over MCP.
Read those two stories next to each other and the message is hard to miss. Make the screen prettier. Also, assume something that can't see the screen is about to become your real user.
Nobody at Apple stood up and said "the UI is dead." They rarely say the quiet part out loud. But it doesn't have to land in the App Review Guidelines as a hard rule to be true.
The front door moved
Here's what I keep coming back to as someone who ships product.
For years the front door to software was a screen. You designed the screen, you argued over the screen, you A/B tested the button on the screen. The screen was the product.
That's shifting under us. The front door is becoming an agent. Siri, Copilot, Claude, whatever your customer happens to be talking to. It reads on their behalf and makes the call on their behalf. Increasingly it acts on their behalf too. It does not care about your hero animation. It cares whether it can understand what you do, and then do it.
I ran into this with our own product. We're a payments company. We built an MCP server for our platform so an agent can look up a payment or kick off a refund without anyone opening the dashboard. Building that changed how I think about the whole thing. In that flow, our carefully designed dashboard wasn't the product. The clean, machine-readable version of it was.
Then everyone rushes to llms.txt and gets it half wrong
The reflex, once this clicks, is to drop an llms.txt on your site and call yourself future-proofed.
I had that instinct too. And clean, structured text genuinely helps. Serving markdown instead of a wall of HTML can cut the tokens an agent burns reading your page by half or more. Some teams report close to 10x. Fewer tokens means the model actually finishes your page instead of bailing halfway through. That part is real.
But be honest about what llms.txt is and isn't. After roughly eighteen months of noise, it's on about one in ten sites, and a big slice of that is Shopify quietly switching it on for every store by default. The bigger problem: the major crawlers from OpenAI, Google and Anthropic mostly don't fetch it. A study across 300,000 domains found it doesn't measurably move your AI citations. If you're adding it as an SEO cheat code, you'll be let down.
Where it actually earns its place today is narrower and more useful. It works as a clean map for coding agents. Cursor, Claude Code, Copilot and the rest read it. Documentation sites get real mileage from it. That's the tell. The people getting value from agent-readable content treat it as plumbing, not marketing.
What I'm telling my team instead
So I'm not chasing llms.txt for citations. I'm asking for something duller and, I think, more durable.
Treat the machine-readable version of every important surface as a real output, not an afterthought. If a page or a screen matters, there should be a clean text representation an agent can consume without scraping your DOM and guessing.
Put a short summary at the top of everything. Two or three lines, plain language, saying what this page is and what you can do here. A summary block helps a human skimming on their phone. It also happens to be the first thing a model reads before deciding whether the rest of your content is worth its context window. You write it once and both readers win.
Expose actions, not just words. Content is turning into table stakes. What an agent actually wants is a verb. "Refund this." "Book that." "Show me last month." App Intents on Apple's side, MCP on ours, a boring documented API underneath. The teams that do well in the agent era won't be the ones with the prettiest content. They'll be the ones whose product can be operated with no human in the room.
And measure the thing that now matters. Can an agent complete your core task, start to finish, without a screenshot? We've started treating that as a first-class test, the same way we treat "can a new engineer run the whole stack in five minutes." If the answer is no, no amount of Liquid Glass is going to save you.
The UI got demoted, not buried
I don't think the UI is dead. I've got designers I'd go to bat for and screens I'm proud of. But it's been demoted. It used to be the product. Now it's one interface among several, and often the one your customer touches least.
Apple just spent a keynote making the screen prettier and, at the same time, making it optional. That's not a contradiction. That's the shift, sitting in one room.
So here's the question I keep putting to my team, and I'll put it to you. If the agent is the new user, what does your product look like to something that can't see?
Top comments (0)