Designing Trustworthy Agentic AI Systems for Autonomous Decision-Making: A Framework for Safety, Explainability, Security, and Accountability

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The rapid development of generative artificial intelligence has led to a new generation of systems known as agentic AI, which can perceive information, reason about objectives, plan tasks, use digital tools, and perform actions with limited human intervention. Unlike conventional AI applications that primarily generate predictions or content, agentic AI systems can operate through multi-step decision-making processes and interact with external environments. This increased autonomy creates significant opportunities in healthcare, education, cybersecurity, finance, public administration, and software engineering, but it also introduces new risks related to safety, security, transparency, privacy, and accountability. This paper proposes a framework for designing trustworthy agentic AI systems based on four central principles: safety, explainability, security, and accountability. The study examines the major challenges associated with autonomous decision-making and presents a conceptual framework integrating human oversight, risk assessment, explainable reasoning, continuous monitoring, secure tool use, auditability, and governance. Analytical tables are used to compare risks, mitigation strategies, trust dimensions, and evaluation criteria. The paper argues that trustworthy agentic AI should not be designed merely as a more capable form of automation; rather, it should be developed as a controlled socio-technical system in which autonomy is proportional to risk and human oversight remains meaningful. The proposed framework can serve as a foundation for future research and practical implementation of responsible autonomous AI systems.