Work has state

Every product demo this year has the same scene. Existing software, a new sidebar, a blinking cursor, ask me anything. We have decided that the interface for intelligence is a text box, mostly because a text box was the easiest thing to ship.

Chat is a good interface for conversation and a poor one for work. Work has state. It has steps that depend on other steps, decisions somebody is accountable for, and outcomes that are wrong in expensive ways. A transcript is the worst available record of all three.

I built Orson in 2018, before agents existed as a category, and none of the hard questions were about conversation. They were which decisions a person will actually hand over, how the interface shows what the thing did and how to stop it mid-action, and what confidence is supposed to look like on a screen when the system is guessing. Those are interaction design problems, the same family as basket logic and age verification on a kiosk, with a probabilistic engine underneath instead of a menu.

An agent acting on your behalf needs terms, and the interface is where the terms get written. A chat box leaves them vague by default. So the products that last will be the ones that design the work first and let conversation be one input among several, the way search stopped being a destination and became a feature.