Trustworthy Agentic AI: A Rapid, PRISMA-Informed Evidence Review of Safety, Explainability, Alignment, and Human Oversight in Autonomous AI Systems

<title>Abstract</title> <p>Agentic artificial intelligence is rapidly emerging as a paradigm in which large language models, retrieval systems, memory modules, software tools, and multi-agent coordination mechanisms are integrated to support autonomous or semi-autonomous decision-making. Unlike conventional AI systems that primarily produce predictions or responses, agentic AI systems can interpret goals, decompose tasks, retrieve external evidence, invoke tools, interact with digital environments, communicate with other agents, and execute multi-step workflows. These capabilities introduce trustworthiness challenges because failures may propagate from generated text into planning, tool calls, workflow execution, external communication, and real-world decision support. This paper presents a rapid, PRISMA-informed evidence review of trustworthy agentic AI across a primary corpus of 33 studies, complemented by one additional structural reference used only to model manuscript organization, focusing on safety, explainability, alignment, human oversight, fairness, privacy, security, evaluation, and governance. The review reports transparent search and screening procedures, a seven-item quality assessment applied to every primary study, and quantitative synthesis of publication trends, source types, and trust-dimension coverage. The paper proposes the T-SAFE taxonomy, organizing trustworthy agentic AI around Transparency, Safety, Alignment, Fairness, and Explainability, and introduces a layered trustworthy agentic AI framework linking environment interaction, memory, retrieval, tool use, reasoning, trust controls, governance, and human oversight. The taxonomy is explicitly positioned against fifteen related academic, industry, and regulatory frameworks published between 2023 and 2026, including four closely related systematic surveys published in 2026, to clarify its specific contribution. Quantitative analysis of the reviewed corpus reveals that Safety dominates current research emphasis (52% of studies) while Fairness receives no primary treatment in any reviewed study, a gap that directly motivates the fifth T-SAFE dimension. The review identifies open challenges in agent safety benchmarking, explainable tool use, secure retrieval-augmented agents, lifecycle alignment monitoring, and human-in-command governance, discusses appropriate manuscript-type positioning for this rapid-review methodology, and explicitly recommends formal multi-database replication as a next step toward a comprehensive systematic literature review.</p>