Guides
Approval boundary
A configurable rule for which AI actions may run, must wait, or never auto-run — tiered by reversibility and blast radius, not a single switch.
Last updated 2026-07-26
An approval boundary is the policy that decides whether an AI action may run on its own, must wait for a human, or is permanently held. Boundaries are category-scoped (not one global “agents on/off” switch) and scale with reversibility and blast radius.
How to think about it
| Kind of action | Typical boundary |
|---|---|
| Read / internal reversible write | Often automatic with a receipt |
| External or higher-blast reversible | Draft freely; commit gated |
| One-way door | Always hold for judgment |
Standing approval can widen a boundary for proven low-risk categories. That is earned autonomy — not blanket trust.