What should a desktop environment be when every session has an agent in it?

VocsCode grew out of this question. Once every session has an agent in it, the desktop app stops being an editor with an assistant bolted on and becomes something else: a harness manager, a permission boundary, a diff reviewer, a cost meter — a cockpit for agents.

But the deeper version of the question is still open. What should persist between sessions? How much should the environment know about what the agent did versus what the human decided? Where does a “project” begin and end when work is delegated but responsibility isn’t?

Current thinking: the unit of work is the session, the unit of trust is the permission mode, and the unit of record is the diff. Whether those three abstractions are enough is what the building is for.