It's 7:40 on a Tuesday and the queue has been filling since I went to bed. Front-line support replies drafted in thirteen languages. The day's numbers reconciled, with two exceptions flagged for a human. This morning's content, written and waiting. Every item sits in one place, with everything I need to judge it on one screen.
This is the part of our operating model people find hardest to picture, so here it is plainly.
Three buttons, three different machines
Approve sends the work on its way and logs a clean approval. Approvals are how an agent earns wider limits — in our lanes, roughly ten clean approvals earn the right to suggest, twenty to thirty earn the right to act, always inside caps. The thresholds aren't ceremony; they're a track record you can inspect any time you doubt it.
Edit is the interesting one. Fix the draft and the correction doesn't vanish — it's fed back and distilled into a rule the agent carries forward. You're not just fixing one reply; you're training the colleague who wrote it. The tier holds: close counts for something, but it doesn't count as clean.
Reject asks you for a reason, in a line, and drops the agent straight back to the bottom tier. One bad call outweighs a streak — deliberately. More on that another week.
Why this is management, compressed
A good manager reads work, judges it, gives feedback that sticks, and extends trust at the pace it's earned. That's exactly what the queue is — with the reading time collapsed, because agents write for a reader: what they did, why, what they want, what it costs. Ten minutes with a coffee covers what would once have been a morning of stand-ups and Slack archaeology.
And underneath all of it sits the audit trail. Every action an agent takes — proposed, approved, edited, rejected, sent — is logged and attributable. The question that kills most AI adoption is "what did it do while I wasn't looking?" The queue's answer: here, read it.
The feeling this replaces is the one every owner knows: work happening somewhere you can't see, done to a standard you can't check. Mornings here start with decisions, not a backlog — and every decision makes the system slightly better at deserving fewer of them.