What five questions define it?#
For each class of work: what may the agent do without asking, when must it stop, who is asked, what evidence travels with the ask, and what happens if nobody answers? Only the first two concern the machine. The last is the one most often skipped, and systems usually discover their answer during an outage at 2am.
What is the goal?#
Maximum useful autonomy inside justified trust boundaries: as much as the evidence supports, with a written reason for where the line sits. Tier by consequence and reversibility, because a one-line change to a trust boundary and five hundred lines of generated test fixtures do not belong in the same tier. Setting levels of automation per class of work replaces the blunt choice between human in the loop and human on the loop.
The line widens through automation follows trust and narrows through a contraction trigger. A person needed only because rollback is missing is system debt.