Дизайнер приравняла AI-агентов к медицинским посредникам, а не к пользователям

This week on Habr, a discussion unfolded about who should be considered AI agents in interface design. Kiara Pellegrini from Nielsen Norman Group proposed simple logic: an agent has goals, interacts with the interface, tries to achieve them. Therefore, it's a user. But a UX designer with eight years of medical experience disagrees: this is not just an imprecision in definition, but a deep misunderstanding of architecture.

Why an Agent Is Not a User

An agent clicks buttons, fills fields, gets results — like a user. But there's a critical difference in motivation and responsibility. A user acts for themselves. An agent acts for you, representing your goals through interpretation. This transforms the entire dynamics of interface interaction. When a designer designs for a user, they create a channel of direct access: a person sees something, understands it, acts. An error is their error of understanding or attention. When an intermediary enters the chain, new points of failure appear. An agent may misread your intention, may forget context, may encounter an exception and not know how to respond.

Medicine Has Known This Question for a Century

A nurse in a hospital is a classic intermediary. A doctor writes a prescription, the nurse executes it, the patient receives medication. Each level can distort information: the doctor writes the dose unclearly, the nurse reads it incorrectly, the patient forgets the instruction. That's why medicine has developed an entire culture of error insurance:

• Double-checking critical procedures
• Standardized forms and protocols instead of free text
• Alert signals when deviations from the norm occur
• Mandatory documentation of each step
• Clear feedback channels between levels

The result: medicine designs not for an ideal intermediary, but for a real one who can make mistakes.

What This Means for Interface Design

If we recognize an agent as an intermediary, not a user, then the designer's task changes. We need interfaces that don't just "work," but work reliably even when the intermediary makes mistakes or doesn't fully understand the context. This means explicit confirmations before critical actions, detailed logging of each step, the ability to roll back operations, and ensuring the agent can read the state at each stage. Today, most interfaces are designed for direct human interaction. When you throw an agent at them, they often break. Because there's no insurance against the intermediary misunderstanding the silent agreement of the design.