Evolving Human-Agent Interaction

Confidence first

We assumed the challenge was navigation. Research said otherwise: people could learn the interface, but they didn’t understand the product. Many saw the assistant and the app around it as two separate things. So I designed the first-run experience to teach one clear mental model, with just enough context to start exploring. Motion helped show where things live, but not why it matters.

Progress, not conversation

In Word, people didn’t want a conversation. They wanted progress. So instead of a blank prompt box, I moved to clear, direct controls, and designed it to ask only when it has to, since a little clarification helps and too much becomes friction. People also wanted a first draft fast, so Word got its own behavior, not a copy of the shared one. Most of all, I treated the document itself as the experience, grounded in your own files and fully editable.

Control, made visible

Memory started as a personalization problem and quickly became a control one: not how to remember more, but how to let people see and delete what’s remembered. I designed it to be transparent but calm, with a notice when something is saved, deletion of single items, a temporary mode, and one place to manage it all. Two calls shaped it.

Not every preference should stick, and transparency has limits, because too little feels scary and too much is noise.