Tag
The author argues that current AI agent safety measures like guardrails and monitoring are insufficient, proposing 'external admission' as a stricter pattern where execution authority is withheld until an external authority explicitly allows high-impact actions.