Trustworthy Agentic AI: A Framework for Critical Systems.

As agentic AI systems, capable of autonomous perception, planning, and multi-step action, are increasingly considered for critical engineering applications, their trustworthiness becomes paramount. This research addresses a gap by defining trustworthiness as a core engineering property, rather than solely focusing on task capability. It introduces a comprehensive trustworthiness model built around five key dimensions: safety, robustness, transparency, accountability, and privacy. The study maps these dimensions onto an agentic assurance workflow, covering everything from perception to audit. It then surveys existing agentic system architectures, potential threats, concrete trust mechanisms, and quantitative metrics relevant for development and evaluation. These principles are examined across diverse, constraint-bound engineering domains such as power systems, autonomous vehicles, high-performance computing, and communication networks. By synthesizing findings across these domains, the research concludes that agentic AI trustworthiness is a unified problem. It outlines a path towards a reusable, cross-domain assurance framework, drawing parallels to the graded certification regimes found in mature safety-critical engineering fields. This aims to provide a structured approach for ensuring that AI agents can be verified, audited, and trusted in high-stakes environments.