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

Technical Summary: Trustworthy Agentic AI

Problem Statement
Agentic artificial intelligence represents a paradigm shift from passive predictive models to autonomous systems capable of interpreting goals, decomposing tasks, retrieving external evidence, invoking tools, and executing multi-step workflows. Unlike conventional AI, where failures are typically limited to model output, agentic AI failures can propagate through planning, tool invocation, and real-world decision support, leading to irreversible actions, data leakage, or malicious execution. The existing literature on trustworthy agentic AI is fragmented across domains such as Large Language Model (LLM) agents, multi-agent systems, explainable AI (XAI), alignment, and retrieval-augmented generation (RAG). This fragmentation creates a lack of consolidated guidance on integrating safety, explainability, alignment, fairness, and human oversight into autonomous decision-action chains. Furthermore, the field is rapidly evolving, with a significant volume of related systematic surveys published in 2026, necessitating a timely, structured review to clarify the current state of the art and identify specific gaps.

Methodology
This study employs a rapid, PRISMA-informed evidence review protocol. The methodology is explicitly distinguished from an exhaustive multi-database systematic literature review; it utilizes a single-reviewer, AI-assisted search strategy supplemented by citation chaining.

• Corpus: The review analyzes a primary corpus of 33 unique studies published between 2022 and 2026. These studies were selected based on English language, publication date, and relevance to agentic AI architectures and trustworthiness dimensions (safety, explainability, alignment, fairness, privacy, security, transparency, oversight, evaluation, and governance).
• Quality Assessment: A seven-item quality assessment rubric was applied to each study, scoring objective clarity, agentic relevance, trustworthiness focus, methodological contribution, limitations, source credibility, and direct relevance on a scale of 0–2 per item (maximum score 14).
• Synthesis: The review combines quantitative synthesis (publication trends, source types, quality scores, and dimension coverage) with qualitative thematic synthesis of architectural patterns and trust risks.
• Positioning: The proposed taxonomy is explicitly compared against fifteen existing academic, industry, and regulatory frameworks (2023–2026), including four systematic surveys published in 2026, to delineate specific contributions.

Key Contributions

1. T-SAFE Taxonomy: The paper proposes the T-SAFE taxonomy, organizing trustworthy agentic AI around five dimensions: Transparency, Safety, Alignment, Fairness, and Explainability. This taxonomy is empirically motivated by the finding that "Fairness" receives no primary treatment in the reviewed corpus, while "Safety" dominates (52% of studies).
2. Layered Framework: A layered trustworthy agentic AI framework is introduced, linking environment interaction, memory, retrieval, tool use, reasoning, trust controls, governance, and human oversight. This framework maps the T-SAFE dimensions to specific architectural components.
3. Quantitative Evidence: The review provides quantitative data on the state of the field, revealing that 73% of the primary corpus was published in 2025 or 2026, and that arXiv preprints constitute the largest source category (42%).
4. Gap Identification: The study identifies critical gaps, including the lack of standardized agent safety benchmarks, insufficient explainability for tool use, security risks in RAG-enabled agents, and the absence of lifecycle alignment monitoring.

Results and Thematic Synthesis

• Architectural Patterns: The review identifies six primary patterns: LLM-based single agents, tool-augmented agents, RAG-enabled agents, multi-agent systems, WebAgents, and self-evolving agents. Each pattern introduces distinct trustworthiness risks, such as hallucinated planning, unsafe tool calls, prompt injection, and cascading errors in multi-agent coordination.
• Trustworthiness Risks: Key risks identified include hallucination (incorrect content becoming a plan), unsafe tool use, prompt injection, weak explainability (lack of rationale for tool selection), misalignment with user intent, and insufficient human oversight.
• Dimension Coverage: Quantitative analysis of the 33 studies shows a heavy skew toward Safety (52%), followed by Alignment and Explainability (15% each), and Transparency (12%). Notably, Fairness was not the primary focus of any study in the corpus, highlighting a significant research gap.
• Benchmark Fragmentation: The review notes that existing safety benchmarks (e.g., Agent-SafetyBench, R-Judge, SafetyBench) evaluate different constructs (task-level safety vs. risk awareness vs. static knowledge) with low inter-benchmark concordance, leading to inconsistent safety conclusions.
• Comparative Positioning: When positioned against 15 existing frameworks, the T-SAFE taxonomy is distinguished by its combination of a five-dimension evaluative taxonomy, explicit mapping to a layered technical architecture, applicability across research and policy without sector-specific narrowing, and grounding in quantitative corpus evidence.

Significance and Claims
The paper positions itself as a rapid evidence review rather than an exhaustive systematic review, acknowledging limitations such as single-reviewer bias and the evolving nature of search terminology. Its primary significance lies in:

• Empirical Motivation: Providing data-driven evidence that current research disproportionately emphasizes safety and security while neglecting fairness, thereby justifying the inclusion of Fairness as a first-class dimension in the T-SAFE taxonomy.
• Structural Clarity: Offering a unified framework that connects technical architectural layers (e.g., retrieval, tool use) with evaluative trust dimensions, addressing the fragmentation in current literature.
• Research Agenda: Explicitly recommending future directions, including the development of standardized agent safety benchmarks, explainable tool-use mechanisms, secure RAG workflows, lifecycle alignment monitoring, and human-in-command governance structures.
• Call for Replication: The author explicitly recommends formal multi-database replication of this review as the necessary next step to transition from a rapid evidence review to a comprehensive, reproducible systematic literature review.

The paper concludes that transforming agentic AI from experimental prototypes into reliable, explainable, fair, and accountable systems requires addressing the identified gaps through standardized evaluation, robust governance, and a balanced focus on all five T-SAFE dimensions.