Writing
Prototype how the agent feels before deciding what it does
Why early, rough, interactive prototypes set the direction for Agent Mode, and for most of the projects on this site.
The most useful thing I did in the early design exploration of Agent Mode was not a spec or a flow diagram. It was a run of quick prototypes of how an autonomous loop would feel inside Word, Excel and PowerPoint. Agent Mode is Copilot with reasoning models and agentic loops behind it, so it can create, analyze and transform content on the fly rather than answer one prompt at a time. That is a big change in what the software does, and it is very hard to reason about on paper. You have to feel it.
So I prototyped interactions, iterated on them rapidly, and shared what I found with the broader team. That rhythm helped establish the direction and informed the research decisions that came after. Looking back, the thing that made it work was the order of operations. Feeling first, then definition.
Here is the argument for that order when the product is an agent.
A conventional feature has a bounded set of states. You can draw them. An agent has a loop: it plans, acts, checks its own work and acts again, and the person either watches this happen or comes back to a result. The interesting design questions live in the loop. How much of the reasoning should the person see? When does the agent pause and ask? What does it feel like to watch a document rewrite itself for twenty seconds? Does the person want to interrupt, and can they? A static mock cannot answer any of those. A prototype with real timing, even a rough one, answers most of them within minutes of putting it in front of someone.
The second argument is about ambiguity. When the capability is new, nobody on the team, including the designer, knows what the right product is. Prototypes are the fastest way to turn a vague sense that something is magical or unsettling into a specific observation that a researcher can test and an engineer can build. Sharing findings early meant the direction was set by evidence rather than by whoever argued best in the room.
I have worked this way since long before agents. On Fidget Digits, a self-tracking device for fidgeting during online lectures, we sketched concepts and discussed them with students, then modelled the designs in Rhino and 3D printed them for evaluation. Feedback from users was taken at every step. Our first prototype focused on functionality over ergonomics, and it was only holding the printed object that told us what had to change. On Makerarm, the teaching mode, where you train the arm by demonstration, was the kind of interaction that only made sense once you could try it. On my own MIDI controller, I tested the layout during rehearsals and fine-tuned mappings, response times and LED feedback based on what my hands did under pressure, not what the diagram said they would do.
Agents raise the stakes on this because the interaction has time in it. Latency, pacing, the moment a result appears, the moment you are asked to confirm: these are the material. They are invisible in a static design file.
A few practical notes from doing it
Prototype the loop, not the screen. The screen is the easy part. Build something that actually takes a few seconds, produces intermediate output and can be interrupted, even if the intelligence behind it is faked.
Iterate in days, not sprints. The value of a prototype at this stage is the number of things it lets you rule out. Speed matters more than fidelity.
Share findings, not opinions. What I brought back to the team was what people did and said with the prototype in front of them. That is what changes research plans and engineering priorities.
Expect the feeling to lead the requirements. On Agent Mode, the aim was that autonomous capability would feel intuitive rather than alarming. That is a felt quality. It becomes a set of requirements only after you have watched people react to it.
The case studies on this site describe the outcomes. The method underneath most of them is the same: make something you can hold or click as early as possible, put it in front of the people it is for, and let what they do set the direction.
More in the writing section, or see the design process these pieces come from.