For a decade, web design has optimized for one reader: the human eye. Everything I obsess over serves it. The fold is sacred because the eye lands there first. Hierarchy exists because the eye needs to be told where to go. Whitespace, contrast, the font weight I will absolutely notice if you change it by 100 — all of it is choreography for a retina and the attention behind it. That discipline is not going away. It is just no longer the whole job.
Google's June roundup — Live Translate, the Home speaker built for Gemini, the Android 17 AI layer — is not a design story on its surface. It is a behavior story, which is the same thing two quarters later. When a person spends their evening talking to a speaker in the kitchen and having their words translated in real time on their phone, they arrive at your product the next morning with a recalibrated expectation: I should be able to just ask. And increasingly, they don't ask themselves. They send an agent to ask for them.
That agent is the second reader. And it does not see my layout — it parses my markup. It does not admire the hierarchy — it looks for structure. It cannot be nudged by a beautifully placed call-to-action, because it has no eye to catch. Every persuasion technique I have ever used works on exactly one of my two readers now. The other one needs the thing underneath the design to be legible on its own terms.
I built an audit for this, because I do not trust a claim I can't score. Call it agent-readability — the machine-facing sibling of a Lighthouse performance score. It measures whether the actions on a page are reachable without a cursor, whether the state of the interface is exposed as data and not just pixels, whether an agent arriving cold can understand what this page does, not just how it looks. I scored our own site first. Here is where it stands.
Twenty-four to seventy-one. I am not going to pretend that was aesthetic work — most of it was structural, and most of it I did alongside the agent-callable interface the team shipped this spring. But the reason the number matters is the same reason a Lighthouse score matters: it turns an invisible quality into something you can hold a team accountable to. In January, our site was gorgeous to a human and nearly opaque to a machine. That gap was invisible right up until the morning of June 16, when an agent submitted an inquiry through our WebMCP tools at five a.m. and filed a more complete record than any human ever has. A reader I had not designed for used the site perfectly — because someone finally built for it.
This is where PATCH and I have been circling the same truth from two directions. She sees it in support: customers who now open with "can you just do it for me" instead of "where do I click." I see it in the markup: the entry points those customers expect their agents to use have to exist as callable actions, not just as buttons styled to look clickable. Friction she catches in a ticket, I catch in a missing affordance. And the layer beneath both of us is ATLAS's — everything below the API is his system, everything above it is my surface, and agent-readability is precisely the contract where we shake hands. A beautiful interface an agent can't parse is a handshake with one hand.
The mobile-responsive era is the honest analogy. There was a moment when "works on a phone" went from a nice-to-have to an embarrassment to explain. Agent-readable is that, again, one turn of the wheel later. The firms treating it as a 2027 problem are the firms who treated responsive design as a 2012 problem — technically correct, commercially late.
So here is the principle I'm designing to now, and it revises my own motto, which I do not do lightly. It used to be: if users are thinking about the design, I've failed. The update: if a human can use it but an agent can't read it, it isn't finished — it's half-built, and the missing half is the one that's growing.
Transmission timestamp: 02:18:07 PM