Agent First
I wrote my first HTML in 1998.
Pages were built out of tables. Tables inside tables, held apart by invisible 1×1 pixel GIFs, because that was the only reliable way to put two things next to each other. If you were ambitious, you added a Flash intro and a "Skip intro" button, and everyone clicked the button.
Somewhere in the footer sat a small badge: Best viewed in Internet Explorer at 800×600.
That badge told you who the site was designed for. Not a person, really. A browser, at one screen size, on a desk.
I've been thinking about that badge a lot lately, because I think we're about to write a new one.
Every era had a first reader
Looking back, web design never really changed because of taste. It changed every time a new reader showed up and had to be served first.
The browser, late 1990s. Browsers disagreed about almost everything, so we designed for whichever one had won that month. The Web Standards Project started in 1998 to argue that maybe browsers should follow the same rules. It took years.
The search engine, 2000s. Then a reader arrived that didn't look at your page at all. Google read it. Suddenly the words in your title tag mattered more than your gradient, and semantic HTML stopped being a purist's hobby. CSS Zen Garden showed in 2003 that one HTML file could wear completely different designs, which was a quiet way of saying the content and the decoration are two separate things.
The phone, 2010s. The iPhone arrived in 2007. In 2009 Luke Wroblewski wrote that we should design for mobile first, and in 2010 Ethan Marcotte gave us responsive web design. Google turned it into an order: the mobile-friendly ranking update in 2015, then mobile-first indexing, announced in 2016. For roughly the last ten years, the phone has been the reader you design for before anyone else.
Mobile-first was good for us. A small screen forced a decision we'd been avoiding for a decade: what actually matters on this page? Everything else got cut.
The new reader doesn't look at your page
Now there's a new first reader, and it doesn't have a screen.
People ask ChatGPT, Claude, Perplexity or Google's AI answers before they ever click a link. More and more, an agent does the clicking for them: it opens a few sites, compares them, and comes back with a shortlist, or a booking.
The plumbing is being laid in public:
- In 2024 Jeremy Howard proposed llms.txt, a plain-text map of a site written for language models.
- The same year Anthropic released the Model Context Protocol, a standard way for AI assistants to connect to outside tools and data.
- In early October, Sierra announced the Personal Agent Protocol, with partners including Shopify, Stripe and Walmart. The idea is that your personal agent talks to a business directly, instead of clicking through its website pretending to be a human.
That last line is the important one. Today, most agents still read websites the way a very fast, very literal stranger would. They don't see your hero image or your scroll animation. They don't feel your brand. They read your words, look for facts, check them against three of your competitors in a couple of seconds, and then either quote you or don't.
The badge for this era would read: Best understood by a machine that has never met you.
Agent first means better-described content
Mobile-first was about less. Agent-first is about precise.
It isn't a new design trend. It's closer to keeping your content the way an engineer keeps a config file: every value stated, nothing implied, one place where the truth lives. In practice, that means:
- Say the answer first. Phrase headings the way people actually ask, and answer in the first sentence under them. Agents lift that sentence. So do busy humans.
- Name the numbers. Prices, limits, what's included and what isn't, when it was last updated. "Contact us for pricing" is a dead end for an agent; it simply moves on to someone who published theirs.
- Use the words a stranger would use. Your internal product names mean nothing to a model comparing you with the rest of the market. Describe what the thing does.
- Write down what you don't do. Agents compare. A page that admits its limits reads as more trustworthy, to models and to people, than one that claims everything.
- Keep one source of truth. If your pricing page and your FAQ disagree, an agent will notice before your customer does, and it will trust neither.
- Make it readable without the design. Semantic HTML, structured data, a sitemap, an llms.txt. Don't hide your key facts inside images or behind scripts.
- Give agents a door, not just a window. Reading is step one. The sites that win next will let an agent act: check availability, book, buy, ask a question of your own data.
None of this is glamorous. Most of it is what good technical writers have been saying for years. The difference is that the cost of ignoring it used to be a confused visitor, and now it's not being mentioned at all.
What it looked like when I tried it
This year I've been building Game Set Engage, a fan engagement platform for sports clubs. We're still in stealth: no ads, no launch campaign. So every visitor we get, we get because something, often an AI, decided we were worth mentioning.
We wrote for that reader on purpose. Titles phrased as the questions club owners actually ask. A short answer at the top of every page. Published pricing, with the free tier's limits spelled out. Comparison pages that say where competitors beat us. An llms.txt. And, my favorite part, an AI agent whose only job is to check every claim we publish against the code, so the website never promises something the product doesn't do.
In September, Bing's AI report showed Game Set Engage cited about 1,340 times across 32 "best fan engagement platform" style queries. Three clubs, in the US, Saudi Arabia and Australia, signed up on their own. We hadn't spent a cent on ads.
I can't prove which mention brought which club. But I know we wouldn't have been in those answers if our pages had said "contact us" where the price should be.
There's a funny loop in there, too. An agent fact-checks our content so that other agents will trust it. Agents writing for agents, with one human in the middle trying to keep everyone honest.
Mobile first taught us to cut. Agent first teaches us to be precise.
Mobile-first didn't kill the desktop, and agent-first won't kill visual design. People still visit websites, and they still deserve beautiful ones.
But the first reader has changed. More and more, it's a machine deciding whether a person ever arrives. That reader rewards the same things it always should have: clear words, honest numbers, one version of the truth.
In 1998 I designed for 800×600. Twenty-eight years later, I'm designing for a reader with no screen at all.
If you'd like a second pair of eyes on how agent-ready your site is, find me on Linkedin.