Agents hand work back, people close it
An agent can now say a task is finished without deciding that it is done, and saved research waits to be read.
An agent working in your workspace could mark a task done. Nobody had approved anything — the work simply disappeared off the list as complete. Research had the opposite problem: an agent could save an entry and there was no state at all for "somebody should read this", so it sat in the list looking identical to everything already reviewed.
Both are one boundary now. Tasks have a review status, and over MCP an agent can reach it and no further: submit_task_for_review hands the work back with an optional note about what it did, and a person closes it in Quiver. The old complete_task tool is gone rather than quietly changed — a tool called "complete" that no longer completes anything is worse than one that isn't there.
Research entries can be marked reviewed, and the dashboard's Needs you section lists the ones nobody has read yet.
Every move is recorded: who handed the work back, who closed it, and through which interface. That is the point of the boundary — it only means something if you can see afterwards that it happened.