Agent Interface
Framework

Help people understand what is possible and express what they want.

The empty prompt is a tax on the user, and most people pay it badly. An interface that starts from a blank line assumes the person already knows the system’s range and how to phrase a request to reach it. A good agentic interface does the opposite: it teaches its own capabilities in the act of being used, and it makes precise expression easy when precision matters. Expressing intent well is a shared task between the person and the surface, not a test the person has to pass alone.

Make the agent’s work visible while it happens.

When software acts on its own over seconds, minutes, or hours, a silent gap is corrosive. The user is left asking whether anything is happening, whether it is the right thing, and whether they should wait or intervene. Showing the work turns waiting into monitoring and lets people catch a wrong turn early. The discipline is to show enough to build confidence without burying the person in every internal step, and to let detail expand on demand rather than by default.

Turn responses into usable interfaces, not just text.

A paragraph describing five options is worse than five buttons. When an agent produces a result, the highest-value form of that result is often something you can see and manipulate, not prose you have to parse and then act on elsewhere. Generated interface closes the gap between the answer and the next action. The layout should serve the task rather than forcing every kind of work through the same scrolling column of text.

Let agents act while people stay informed and in command, during the work rather than only after it.

Autonomy without control is not a feature, it is a liability. As agents do more, the interface has to keep the human able to steer, and to place that steering inside the flow of execution instead of presenting a finished result to accept or reject. Control is not a wall at the end of the process; it is a set of handholds along the way. The bar rises with the stakes: the more sensitive, expensive, or irreversible an action, the more explicit the human’s say-so must be.

The interface must earn belief rather than assume it.

Trust is not a tone of voice, and confident prose is not evidence. An agent that sounds sure while being wrong is more dangerous than one that is transparent about what it knows and how it knows it. The interface is where trust is calibrated, up when the system deserves it and down when it does not. That calibration depends on the surface telling the truth about its own certainty, its sources, and its memory.

Assume the agent will fail, and make failure survivable.

Agentic systems are probabilistic, which means they will be wrong, blocked, or interrupted, and the only question is whether the interface was designed for that or caught off guard by it. Most trust is lost not at the first error but at the first error that could not be understood, corrected, or undone. A surface that plans for failure keeps a bad step from becoming a dead end. Recovery is a first-class part of the design, not an edge case bolted on later.

Work should produce durable outputs, not evaporate into a transcript.

The thinking can happen in the conversation, but the product of the work should live somewhere stable. When results exist only as messages scrolled past, the work is effectively lost the moment the session ends. Artifacts are not only an endpoint: they accumulate, get revised, and are handed off, so the interface should treat them as living objects from the start. The test is simple: if the conversation disappeared, would the work still be there.

Give agentic work a persistent, coherent place that fits the product it lives in.

A single task is not a product. As agentic work accumulates across sessions, threads, and outputs, it needs somewhere to live that holds together and belongs to its surroundings. An agent grafted onto a product as a floating box in the corner will always feel like a guest. The interface should give the work continuity over time and coherence with the brand, domain, and surfaces around it, and it should learn from how the person corrects and steers it.