The most powerful AI agents can plan trips, write code, and run complex workflows and then fail on something simple like a CAPTCHA.
Not because they’re not smart. Because the web wasn’t built for them.
Today we have agents that can think, operating in a world designed for humans to click and scroll.
The assumption nobody wrote down
Modern software is designed around one default user: a person sitting at a screen, using eyes and hands to find the right button.
That assumption became so normal it disappeared.
Agents aren’t that user.
An agent doesn’t “browse” the way a person does. It doesn’t notice visual cues or hunt through menus. It needs:
- clear inputs
- clear outputs
- a predictable interface
Everything else — layout, color, marketing copy — is extra work and extra failure points.
The wrong instinct
When builders realize this, the common response is to add AI on top of the existing UI:
- a chat widget
- an “Ask AI” button
- an AI summary panel
That helps, but it misses the real shift.
It’s like adding a phone line to a restaurant because customers now order via delivery apps. The phone isn’t the system. The kitchen is.
The shift is that software now has a second kind of user. Agents already call APIs, move data, and trigger workflows across tools. They don’t need prettier dashboards — they need reliable operations.
So the question isn’t only “is this easy for a human?” anymore.
It’s also: “can an agent reliably use it?”
What friction looks like for an agent
A simple example: the agent needs numbers from a SaaS dashboard.
To get them, it often has to:
- log in (sessions, cookies, 2FA)
- load a JavaScript-heavy app that returns mostly empty HTML
- run a real browser to render the page
- scrape the DOM and hope the layout hasn’t changed
All of that just to reach data that already exists as structured records somewhere.
The UI is great for humans. For agents, it’s mostly latency and fragility.
What “building for agents” actually means
It doesn’t mean removing the human UI. Keep it for people.
It means adding a second interface designed for agents:
- APIs instead of UI-only workflows
- structured data instead of rendered pages
- stable, composable operations instead of giant navigation trees
- consistent contracts instead of personalization
Humans like experiences that adapt. Agents need systems that behave the same way every time.
Add a second door
The best builders aren’t simplifying their products. They’re adding a layer.
They ask: if an agent needed to do everything a human can do in our product, what’s the clean path?
That looks like:
- an API next to the dashboard
- a published spec
- useful, structured errors (code + reason + suggested fix)
- machine-readable outputs
- tools that describe their own capabilities
This isn’t exotic. It’s the same discipline that made good developer APIs — applied to a new kind of user.
The question that’s changing
Jakob Nielsen helped define usability for human users. He recently argued that traditional UI design becomes obsolete when agents are the primary users of digital services.
Whether or not that’s fully true, one thing is: the definition of “usable” is expanding.
Builders who see it aren’t abandoning UX. They’re extending it.
Can a human use this easily?
And can an agent reliably operate it?
Most software can’t answer the second question yet. The ones that can will define what comes next.