The first thing we write on a new workflow is not a prompt. It is a sentence describing what the agent must never finish on its own. Everything else is negotiable; that sentence is what makes the rest of it safe to run.
Why teams resist without one
An operations lead who cannot say which cases still reach them has no way to plan a week. That uncertainty reads as distrust of the model, but it is really distrust of the boundary.
Write the boundary down and the conversation changes within a day. Teams stop arguing about whether the agent is good enough and start asking for more of the queue.
What the rule looks like
One page, in the client’s language, not ours. It names the categories the agent may close, the categories it may draft but not send, and the categories it must never touch.
- Close alone. Anything the eval suite covers at the agreed pass rate.
- Draft only. Anything unusual but recoverable, with the draft attached for a person.
- Never. Billing changes, legal complaints, anything where being wrong costs more than being slow.
Autonomy is earned in public
Categories move from draft-only to close-alone when the evals say so, in the Thursday review, with the run log open. Nobody has to take our word for it, which is the point.
The result
On the helpdesk we ran this on, eighty-four percent of tickets now close without a person opening them. That number came from writing down the sixteen percent first.
002
/ Keep reading
From the journal