Three surfaces, three user shapes
Three surfaces, deliberately different
An assistant from a single provider now usually appears as three distinct surfaces. A visual generation surface that produces slides, prototypes, one-pagers, and images from natural-language briefs. A desktop assistant aimed at non-technical users that orchestrates files, applications, and external connectors through a familiar window. A terminal agent that operates directly on a file system and is most powerful in the hands of someone comfortable with the command line. Each was designed deliberately for a different kind of user. The temptation to use one for everything is the productivity equivalent of carpentry with a single tool.
The visual surface — where deliverables live
The visual surface is the right home for any piece of work whose final form is something a reader will look at rather than read closely. Decks, prototypes, marketing pages, infographics, one-pagers, dashboards in their first sketched form. The strength is fluency with layout and image; the weakness is that the surface is not where complex reasoning happens. A reasonable cadence is to draft structure and content elsewhere, then bring it to the visual surface for the final form. Doing it the other way round tends to produce attractive artefacts that have not been thought through.
The desktop assistant — where coordination happens
A desktop assistant is the right surface for cross-functional work that touches many tools — calendar, email, documents, project trackers, customer-record systems — and that needs to be approachable for someone who would not voluntarily open a terminal. The strength is breadth; it can pull data from many places, write to many places, and run things on a schedule. The weakness is that it is less powerful per task than a deeper specialist would be. The honest position is that desktop assistants are the surface through which most non-technical workers will end up doing day-to-day AI work; serious operators add a deeper tool alongside, rather than instead.
The terminal agent — where compounding lives
A terminal agent is the right surface for anyone whose work is recognisably a system — code, structured notes, scripts, datasets, generation pipelines. The window is unfriendly at first and almost embarrassingly powerful once it stops being. The reason productivity professionals end up here is that the local-folder pattern compounds. Everything the worker builds — instructions, skills, scripts, knowledge — is files; the agent reads them every time; and the system gets better in a way the chat surfaces cannot match. The right time to add it is when one's working habits have started to feel constrained by what a desktop assistant can store about them.
How most working setups end up arranged
A common steady state, observed in working practice, has all three surfaces in play with clear roles. The terminal agent runs the day's deep work and owns the local folder of skills and knowledge. The desktop assistant handles cross-tool coordination, the daily briefing, and the work that touches systems an individual contributor cannot trivially script. The visual surface produces the deliverables. The boundaries are not absolute and shift with task and mood; the discipline is that no single surface tries to do all three jobs, because each of them was designed not to.
If you had to do all of next week's work through one interface — a chat box, a desktop assistant, or a terminal — which would you pick, and which kinds of work would suffer most?