Reliability
Approval gate
A checkpoint where an AI-proposed action requires human sign-off before it executes.
The critical detail is where the gate lives. A prompt saying "always confirm before sending" is a suggestion the model can drift past. A gate enforced in your backend cannot be argued with.
Sensible practice is to grade actions by consequence:
- Low — let the agent proceed (looking something up, drafting text).
- Medium — proceed, but log it and make it reversible.
- High — the agent proposes, a person applies. Customs declarations, refunds, anything a customer sees.
Deliberately building the high-consequence tools as read-only, so they can only ever propose, is a design decision rather than a limitation.