The Human Escalation Mechanism (HEM) for Agentic AI Systems

AI agents operating autonomously encounter situations where their authorized action space is insufficient for the task at hand, where policy mandates human judgment regardless of agent capability, or where the agent itself assesses that human oversight is required. No existing standard specifies the kernel-level contract that governs what happens in these moments: what triggers escalation, what a human principal may decide, how the system behaves while awaiting that decision, and what happens when the designated human is unreachable. This document defines the Human Escalation Mechanism (HEM): a normative protocol in which a governing OS kernel places an agent session into a formally defined HEM_PENDING state, routes a structured escalation request to one or more designated human principals, enforces a prohibition on state transitions until a human decision is received, and processes five defined human decision types. HEM is a first-class kernel session state, not an application-layer confirmation protocol. This document also defines the Policy Rationale Declaration (PRD), which links Cedar policies that route to HEM with machine-readable rationale, and the Decision Rationale Record (DRR), which captures the human principal's reasoning for high-consequence decisions. HEM provides the technical specification for human oversight of agentic AI systems required by EU AI Act Article 14.