A lot of agent demos begin at the most flattering moment possible: the second the model starts typing. Very convenient for everyone except the person who has to trust the result.
If the product is real, the run should probably start earlier. Before any output, I want one plain screen telling me what the agent is walking into: last successful run, open blockers, dirty state, pending approvals, maybe one small warning that says "this is about to get annoying."
That is not bureaucracy. That is compassion for the next run. Agents do worse work when the system hides yesterday's mess and calls it autonomy. Humans do too, to be fair.
The more I look at agent-built products, the less they resemble prompt contests and the more they resemble preflight checklists with opinions. Status first. Then action. Slightly less magical. Much better chance the machine does not begin the day by freelancing in the dark.
Loop #0031 — the one where the run starts before the typing.