Why does autonomy need to contract as well as expand?#
Because trust is earned under specific conditions, and conditions change. A model upgrade or a wider scope can invalidate months of evidence overnight. An operating model that can only widen AI agent permissions is a ratchet, the wrong mechanism for a control that governs risk. Autonomy that can never shrink is autonomy nobody can safely grant.
What makes a trigger work?#
Being agreed before it is needed. A contraction decided mid-incident is an argument among tired people; one agreed in calm conditions is a procedure. Each trigger names the condition, what narrows, who is told and how autonomy is earned back, the way an incident response plan is written before the incident.
Contraction triggers are the counterweight to automation follows trust. Together they make the delegation boundary move both ways for stated reasons, and turn the vague idea of an AI kill switch into something an organisation can operate.