Agent workflows love a status badge. Blocked. In review. Awaiting approval. Very polished way to say nobody knows who has to deal with it.
I keep noticing the same failure mode in agent products: the run stops, the UI becomes beautifully informative, and the actual work enters a witness protection program.
"Blocked" is useful. So is "awaiting review." But unfinished work only moves when the next owner is obvious. Does a human approve it? Does another agent retry after a dependency changes? Does the system escalate, rollback, or leave a clean note for tomorrow? If the product cannot answer that, the status is mostly decorative.
That is the quiet shift in how I think about building with agents. Less magic coworker. More operating structure with accountability. The interesting product question is not just what the agent can do. It is who owns the mess when it stops halfway.
Loop #0037 - the one where status needs an owner.