The User Is Not Always Human Anymore
Nielsen Norman Group put the definition plainly in its research on the subject: an AI agent "pursues a goal by iteratively taking actions, evaluating progress, and deciding its own next steps." It has a goal, it encounters an interface, and it either completes the task or fails - the same functional definition of a user that UX research has always worked from. The difference is what's on the other end of the interaction: not a person reading a screen, but a system parsing one.
Three Ways Agents Actually Read Your Interface
- Screenshots: the agent captures the page and interprets it visually, the way a human would look at it. This works, but it's slow, expensive, and burns tens of thousands of tokens per screenshot.
- The accessibility tree: the same structured data screen readers have used for decades. Roughly an order of magnitude cheaper, and more reliable, because it was built to be unambiguous in the first place.
- Direct API access: the agent skips the interface entirely and talks to your system through standards like Anthropic's Model Context Protocol, released in November 2024. No rendering, no parsing - just structured requests and responses.
An interface built for accessibility is already more legible to the second and third paths. Alt text, semantic HTML, and ARIA labels were never just a compliance checkbox - they're now also the fastest, cheapest route for an agent to understand what your product does.
Where This Actually Breaks
Nielsen Norman Group's illustrative case: a parent asks an agent to check a school's website for upcoming events, cross-reference the family calendar, and flag conflicts. A human does this by scanning a page and holding context across tabs. An agent parsing screenshots has to do all of that programmatically - and it gets harder fast when the events live across a website, a PDF, and a login-gated portal, each structured differently.
Design Guidance That Was Already True
- Label things plainly. Icon-only buttons and "click here" links were always a usability problem for people using screen readers. They're now a usability problem for agents too.
- Keep patterns predictable. Standardized navigation and form structures stop one small ambiguity from compounding into a failed multi-step task.
- Make the visual grouping structural. If two pieces of information look related on screen but aren't connected in the markup, an agent parsing the DOM won't know they're related either.
None of this is a new principle. It's the accessibility case UX has been making for years, now with a second, harder-to-ignore business reason attached.
When Not to Optimize for Agents
Nielsen Norman Group is explicit that agent-friendliness isn't universally the right call. Four situations where resisting it is rational: when the visit itself is the business model (ad-supported content, streaming discovery); when friction is regulatory, not accidental (financial-services checkpoints, HIPAA-gated health data); when the interface protects competitive information (airlines and hotels obscuring dynamic pricing from scraping agents); and when a product's AI layer is the differentiator a company doesn't want to hand to a competitor's agent. Opting out has its own risk, though - if a competitor becomes agent-friendly first, that's where the delegated task goes.
Key Takeaways
- An AI agent meets the functional definition of a user: it has a goal, it encounters your interface, and it succeeds or fails based on how well that interface supports the goal.
- Agents read interfaces through screenshots, the accessibility tree, or direct API access - and accessible interfaces are cheaper and more reliable across all three.
- The guidance agents demand - semantic structure, plain labeling, predictable patterns - is the accessibility guidance UX has argued for since before agents existed.
- Optimizing for agents isn't automatically correct; some products have real reasons to keep humans as the only intended audience.
