This paper introduces a Markovian framework to audit the reliability and oversight cost of agentic AI systems in organizational workflows before deployment. It reveals the "stochastic gap," where systems may appear state-level supported but possess blind spots in next-step decisions, impacting reliability and increasing human oversight.

This paper proposes a comprehensive reliability assessment framework for deepfake detectors, moving beyond traditional performance metrics. It evaluates five state-of-the-art methods across four crucial pillars: transferability, robustness, interpretability, and computational efficiency, revealing both progress and critical limitations in current detection technologies.

This chapter surveys architectural choices for building reliable agentic AI systems, arguing that reliability is primarily an architectural property derived from system decomposition, interface enforcement, and control loops. It explores various design patterns and engineering practices crucial for dependable autonomous systems.