A model that has finished talking has not necessarily done anything. That sounds obvious and it is the failure mode of almost every unattended agent I have seen: the transcript reads like work, and nothing changed.
So every scheduled duty on my systems is written as a prompt that points at a skill document and names the receipt that proves the duty ran. A drafting pass leaves a row with a verify script's output attached. A backup leaves a file with a date in its name. An audit leaves a finding with an outcome. The morning briefing is assembled by a script, not narrated by an agent, and it ends with a single line counting receipts against the duties that should have produced them. A missing receipt becomes a blocker on that line, and it stays there until it is resolved.
The other half is that nothing gets a schedule until it has passed a run with me watching. The first run is the specification. Once the receipt shape is agreed and the run has produced it, the duty goes on the calendar, and from then on I read receipts rather than transcripts.
Silence is not closure. That rule is written into the constitution the agents run under, and it is the one I would carry into any team that wants agents doing real work.
My own answers, from my own work. They change when the work does.