// System Personas and UCD
An AI agent doesn’t care about your brand guidelines. It cares whether you’re useful. And how efficient.
User-centered design gets a new kind of user. Not a persona you write for a user, a system standing in the queue as one.
Also: screen readers, search spiders, and a word the web has been using since the nineties without noticing what it meant.
The new user
Design has always had a cast of users you build personas for. The busy parent. The nervous first timer. The power user who skips onboarding entirely. Add one more to the list. The agent, booking your restaurant table, comparing your prices, filling in your form, reading your terms and conditions so a person doesn’t have to.
Two different things hide under that one word, though. There’s the AI baked into your own product, the copilot with its own private hooks into your backend. And there’s the agent that turns up from outside, working for someone else. Build for the first and you don’t need the interface at all. Give it its own API and let it skip the page entirely. The second doesn’t get that shortcut. It has to read your actual page, the one you built for a person. No signup, no key, and no control over what it does once it’s there.
I wrote about system personas before, documents that describe what a piece of software wants and how it should behave. This is the other side of that. The agent isn’t just running on a persona anymore. It’s arriving as one. Your product now has to be user-centered for something that has never seen your homepage’s hero animation and never will.
What it actually wants
Here’s the uncomfortable bit for anyone who’s spent a career on brand. To an agent, your brand isn’t your logo. Isn’t your colours, your photography, your tone of voice, your design system. Your brand is what you do, how you do it, and how easily it can interact with you to get what it wants.
That’s it. That’s the whole persona.
It’s accessibility
Which means the discipline that already solved this problem is accessibility. Structure content so a screen reader can navigate it, and you’ve mostly structured it so an agent can too. Clear headings. Real labels, not colour and placement doing the labelling for you. Buttons that say what they do, not ones that only look like buttons because of a shadow and a hover state. Content that exists in the markup, not content painted onto a canvas by JavaScript on the assumption a human is watching.
The overlap isn’t a coincidence. Both users have the same complaint. They can’t see your branding. They can only work with what the structure actually gives them. Design for one and you’ve mostly designed for the other.
This is where the split from earlier actually matters. Your own AI was never blind, it was handed the data directly. The outside agent doesn’t get handed anything. It has to perceive the page the way a screen reader does, blind to your visual design, working from nothing but the markup.
That’s not the whole job. There’s still structured data, sitemaps built for machines instead of search engines, the plumbing that tells an agent what’s actually available to it. But readability is table stakes. Get that wrong and none of the rest matters.
I’ve said something like this before, about a shoe with a collapsible heel.
The map
Accessibility gets you into the page. Information architecture is what tells you where you are once you’re in it, and where everything else lives in relation to it.
It’s an old discipline. Taxonomies, hierarchies, labelling systems, sitemaps, the actual shape of an information space. Peter Morville and Louis Rosenfeld wrote the book on it in 1998, literally, the polar bear book, back when most of the web’s problem was navigation rather than decoration. A lot of it comes from library science. Cataloguing existed long before websites did.
Most product designers today have never studied it. The discipline got quietly folded into UX at some point and then mostly discarded, because feeds and search bars made navigation feel like someone else’s problem. Why build a sitemap when a search box finds it for you.
An agent doesn’t get to skip that step. It has to reason about the shape of your product to complete a task. What’s a category, what’s a filter, what’s a synonym for what, what sits nested under what. Bad information architecture doesn’t just confuse humans anymore. It makes your product illegible to the thing trying to act on a human’s behalf.
Same split, different discipline. Your own AI doesn’t need a map, it already has the coordinates. The outside agent has to actually find its way, category by category, the same as anyone walking into a building for the first time with no floor plan and no directory.
Good IA was always about one thing. Can someone find their way around without you standing next to them explaining it. Turns out the someone doesn’t have to be human.
No patience
A human can survive bad information architecture. Confusion, some backtracking, a support ticket if it’s bad enough, but people generally muddle through. We’re forgiving. We pattern match from every other product we’ve ever used, guess at an ambiguous label, learn a product’s quirks over a few visits and adapt to them without even noticing we’re doing it.
An agent doesn’t get all of that. It’s seen more interfaces than any single human ever will, so it’s good at convention, standard buttons, standard forms, standard menus. What it doesn’t get is a memory of your product specifically, the quirks it worked out last time, the label it eventually learned meant something odd. Every session with your product starts as if it were the first. Give it an ambiguous label that only makes sense once you already know your product, and it doesn’t shrug and guess sensibly the way a returning customer would. It either fails outright or, worse, confidently does the wrong thing.
Steve Krug’s classic is called Don’t Make Me Think. Aimed at humans, that title was always a bit of a metaphor. People can think, they just resent having to. For an agent it stops being a figure of speech. There’s no reservoir of goodwill to draw on if your structure makes it think. It doesn’t get mildly annoyed and push through anyway. It just fails the task.
Someone should write Don’t Make Me Think for the agent era. The rules would be the same ones Krug wrote in 2000. The stakes for ignoring them just went up.
We already had a word for this
The amusing part. None of this needed new vocabulary. The web has called this a user agent since HTTP grew a header for it. A browser is a user agent. A search engine spider is a user agent. Every request that’s ever hit a server has arrived with a field announcing exactly what kind of agent it is.
SEO was always accessibility for spiders. Nobody called it that at the time, but that’s what it was. Structured headings, sensible markup, content a crawler could actually parse instead of a slideshow it had to guess at. The industry’s rebranding it now, answer engine optimisation, AI visibility, whatever the slide deck calls it this quarter. Same job. Newer, more literal user.
The same lesson twice
The accessibility community has been saying all of this for decades, largely ignored by anyone who didn’t legally have to listen. Semantic structure. Plain language. Content that doesn’t depend on how something looks to make sense. They were right for a person using a screen reader. Turns out they were also right for that person’s agent, doing the same job on their behalf.
It took a non-human user to make the rest of the industry pay attention to a lesson that was always about people first.
I'd love to tell you more.
Previous Journal Post
// Thanet Beach Conditions
I wanted to know if the sea’s actually worth the drive before I get there. Air temp, wind, rain chance, water temp, waves, tide times, daylight and UV, air quality and pollen, live sewage status and last season’s water quality rating, for every beach on the Thanet coast, in one page, with a one-sentence verdict at the top so you don’t have to read the rest unless you want to.