The canvas is a third participant
What classic conversation analysis reveals when you apply it to working with an agent in a canvas tool: the artifact joins the conversation as a third party.
Classic conversation analysis assumes two parties passing a floor back and forth, one turn-constructional unit at a time. Sit with a transcript of real work in a canvas tool like Claude Code and that picture stops holding.
The turn is no longer the utterance. One “turn” by the agent can be forty silent actions in a row: read, edit, render, check, fix. The human cedes the floor for the whole stretch and only re-enters at completion. Turn-taking has become floor-dispatch. The design question shifts from “whose turn is it” to “when does the agent hand the floor back.”
And there is a third party in the room: the artifact. The second half of every request-response pair is not really discharged by talk. It is discharged by a change in the world, a page that now renders, a commit that now exists. So the human stops checking the words and starts checking the canvas. “Not there yet.” “I don’t see the fonts change.” “Have you pushed?” Trouble can now originate in three places, not two: in the talk, in the agent, or in the world. A surprising share of repair is aimed at that third party.
As the canvas carries more of the shared meaning, the talk shrinks. Human turns collapse toward pointing and assessment: “here”, “this one”, “not there yet”, “I love it”. You can be telegraphic because the artifact holds the referent for you.
Which leaves the open question. If the artifact is a participant, conversation design is no longer the design of an exchange between two parties. It is the design of a three-way system: human, agent, and the thing being made. What does turn-taking even mean when one of the participants cannot speak, only change?