Agents operate subsurface. Humans work at surface level. Agent-first is real, but we optimized for use long enough. Now it’s time to optimize for service.
She walks around the corner, white coat and bright yellow clogs. “Hello Mr. Morrison, how are you doing today?”
While having her daily catch-up with the patient, she grabs the phone from her pocket and starts typing.
“Hello?” Mr. Morrison grabs her hand to get her attention.
She quickly presses send and fakes a smile to get back into the conversation.
Since IT deployed the new med-tech app, the doctor talks to twice as many patients: the actual person and their digital twin. A command line now lets her ask the system to run routine tasks just by talking to the model.
But the actual job got harder. They terminated her easy-to-understand patient chart with the insights that told her what a patient needed. Everything sits one command away, but she’s having two conversations at once.
Nobody did it deliberately. But somebody made an interface decision without asking what her job actually is, and that made it worse.
Is the GUI Dead?
I don’t mind that we move toward command-line interface (CLI) and agentic workflows. I do mind that we often move ahead too fast and forget who we are building for, skip design and terminate the graphical user interface (GUI).
Design experts like Don Norman, Luke Wroblewski and Jakob Nielsen come to a pretty consistent conclusion. Design principles like affordance, signifiers, feedback and conceptual models came from doors and elevators, long before we used them for screens. They helped us design for computers. They’ll help us design through this transformation too.
Nielsen puts it plainly: UX will not die, but will metamorphose. The way we interact and what we interact with will change.
And it’s changing fast. In the first quarter of 2026, six major repositories launched ways to take a GUI that already exists and wrap it in a structured CLI that agents can use. All six crossed 4,000 stars within days. CLI-Anything explains: today’s software serves humans; tomorrow’s users will be agents.
The truth is, if your software has no proper API, no MCP server, no CLI, it’s effectively invisible in an agentic workflow. An agent can’t reach it, so your software might not be as usable anymore.
The way mobile-first reshaped what we built, agent-first will reshape it again.
Do We Need a Manual?
Think about automatic car transmission.
In the early years, drivers still needed the clutch. We built systems under the hood that could shift for us, but it took many iterations and attempts to win the trust to make it the default.
At some point, we stopped thinking about the clutch at all. Now shifting is just prompting our intent. Park, Reverse, Neutral or Drive. Four states that tell you exactly what the automated system will do, so you can rely on it blindly.
But it took nearly two decades to become the main type of transmission, and we still use the shifter to set it in the right direction.
My frustration with chatbots is that they often don’t indicate what they can do. No P/R/N/D. At a glance, the system feels powerful, and then you’re three sessions deep in a negotiation with a machine about what you actually meant. You need prompt training, or MCP know-how, before the tool does what you want.
It feels like they removed the clutch and the shifter overnight and now just expect you to trust the system, like stepping into a robotaxi. Going from a manual to a robotaxi is a big leap. Outside the early adopters, it will take many in-betweens before we build the trust to give up on graphical interfaces entirely.
Surface and Subsurface
Agents operate subsurface. That’s where the APIs, the schemas, the deterministic output and the guardrails belong.
Humans operate at surface level. Seeing, comparing, pointing, judging, deciding. That requires human-centered design, which is often skipped.
The failure in that hospital ward was handing a doctor the subsurface instead of designing a new surface with a better engine underneath.
Put that same chart on a surface by the bed. Or put it on a tablet so she can turn toward Mr. Morrison, click the GUI, let the agent help, and show the patient what she’s looking at. Allow the doctor to walk, talk, point and interact. Let the system do the rest. Service through surface.
Closure
I think we need to build products with great subsurface functions. But we should remind ourselves that the surface level will still be about interactions between humans and computers in real-life scenarios. We optimized for use for years. Now it’s time to optimize for service.
So use the terminal when you already know what you want executed, and terminate the interface that only exists for manual labor like input forms.
Everything in between is a deliberate design decision, made in the context of somebody’s actual job to be done.
To give Mr. Morrison the best care, we need to care about the doctor.
Support with the subsurface. Serve with the surface.
