Human-centered design emerged in the 1980s as a counterweight to technology-first thinking. The principle was simple: build systems around how people actually think and behave, not around what’s easiest for engineers to build.
Forty years later, that principle faces its most interesting challenge yet. When AI agents can perceive context, reason about goals, and take actions—often faster than any human could—what does “human-centered” even mean?
The Inversion Problem
Traditional HCI placed the human in the loop by design. Every meaningful action required a human gesture: a click, a tap, a voice command. The interface was the membrane between human intent and machine execution.
Agentic systems dissolve that membrane. An agent given a goal can execute hundreds of sub-actions without ever surfacing a decision to the human. This is what makes them powerful and what makes the design challenge so interesting.
Instead of designing an interface for a human to control a machine, we’re designing systems where machines control workflows, humans set direction, provide judgment at key moments, and evaluate the products.
Principles for Agentic UX
Based on early experience deploying these systems, we’re developing a working set of principles:
1. Make agency easy and observable. Humans need to understand what an agent is doing and why. Invisible agents erode trust. Show the work.
2. Design for meaningful interruption. Not every decision needs human approval, but the right decisions do. The design challenge is identifying those moments and surfacing them gracefully.
3. Preserve human override. A well-designed agentic system always gives humans a clear path to pause, redirect, or countermand. This isn’t just a safety feature—it’s what makes delegation feel comfortable.
4. Account for mistakes. Agents will err. Design for recovery, not just for success. Make it easy to understand what went wrong and support undo, recovery, and graceful adjustment.
5. Build trust incrementally. Humans extend autonomy to systems they’ve learned to trust. Start with narrow delegation and expand as confidence grows.
These principles are hypotheses to be tested.