This article argues that AI agents often execute actions based on incomplete instructions by guessing missing information, a problem termed "pre-execution confirmation failure." It proposes a runtime-enforced structure that requires verifying knowns and unknowns before any action is taken.
- The core issue is not the rule "if unsure, ask," but the lack of structural enforcement preventing AI from filling gaps.
- Level 1 failures involve AI guessing and filling gaps, while Level 3 failures occur when actions cannot be confirmed from user words.
- The proposal includes a 12-slide mechanism designed for AI providers and agent platform developers to implement pre-execution checks.
- This approach aims to stop agents from creating accidents by forcing them to pause and verify conditions before execution.
The authors consider this important because it shifts the responsibility of verification from intention to structure, ensuring that AI does not execute actions when required information is unverified.