One Gate Is Not Enough: Composing Stateful Pre-Action Controls for Agentic AI

This research introduces a formal framework for understanding and managing pre-action controls in agentic AI systems, specifically addressing situations where multiple controls (e.g., authority, resource, evidence gates) govern a single action. The core finding is the concept of 'remediation-induced control coupling,' where the remediation by one control can inadvertently alter the conditions for another, potentially invalidating prior judgments. The paper proposes a protocol to restore per-action soundness in these complex interactions.

Why it matters

The increasing deployment of agentic AI systems necessitates robust and reliable control mechanisms, especially for consequential actions. Understanding how multiple pre-action controls interact and potentially interfere is critical for ensuring system safety, compliance, and trustworthiness. This research highlights a fundamental challenge in AI governance and offers a theoretical basis for developing more secure and predictable AI operations.

Key insights

• Agentic AI systems often require multiple pre-action controls (e.g., authority, resource, evidence gates) to govern consequential actions.
• A new concept, 'remediation-induced control coupling,' is identified, where one control's remediation can impact another control's evaluation.
• This coupling can invalidate earlier judgments made by other controls, creating complex interdependencies.
• A 'remediate-and-regate' protocol is proposed to restore per-action soundness in bounded, idempotent settings.
• It is shown that implemented remediation operators, such as evidence substitution and resource-budget downroute, do not commute, meaning the order of application matters.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.18360