Something quiet happened at the end of last year. Agents started deploying agents. The feedback loop compressed.
What used to take weeks — a sprint, a review, a staging environment, a deploy — is beginning to collapse into hours. Not because humans got faster, but because some of the humans in that loop are no longer human.
A New Kind of User
When we design interfaces, we have always imagined a person on the other end. Someone with fingers and cognitive load and impatience. A person who might misread a label or miss a button in their peripheral vision.
Agents do not have peripheral vision. They parse what is visible, structured, and consistent. The ambiguities we design around — the forgiveness built into large tap targets, the signposting for confused first-time users — are invisible to them. They either find the action or they don't.
This is not an argument against human-centered design. It is an argument for extending the definition of "human-centered" to include the tools humans use to act on their behalf.
The Interface Layer Is Not Going Away
Some have argued that agents will eventually bypass interfaces entirely — that API-first infrastructure makes UI obsolete. This is partly true and mostly wrong.
Interfaces persist because they are how humans maintain legibility over automated systems. The dashboard for an AI agent is not for the agent — it is for the person reviewing what the agent did. The audit log, the approval flow, the pause button: these are human affordances layered over automated action.
The interface layer is not going away. It is becoming a layer of governance more than a layer of input.
What This Changes
If agents are increasingly the primary users of infrastructure, and humans are increasingly the auditors of that infrastructure, then interface design has two distinct jobs:
- Make agent action legible: clear state, predictable behavior, structured outputs
- Make human oversight possible: trails, summaries, intervention points
These are not at odds. But they require a designer to hold both users in mind simultaneously — the one who acts and the one who reviews.
The most interesting interfaces being built right now are not for people or for agents. They are for the conversation between them.