PRISMA 2020 Registration — Agentic Artificial Intelligence: Foundations, Architectures, Critical Evaluation, and Safety Governance — A Systematic Review of Peer-Reviewed Literature (2015–2025)

Agentic Artificial Intelligence — encompassing large language model-driven systems capable of autonomous goal pursuit, multi-step planning, dynamic tool orchestration, and adaptive self-improvement — represents a fundamental paradigm shift in applied AI. Despite exponential publication growth, the field lacks consensus on architectural design principles, evaluation standards, and safety governance frameworks applicable to deployment. This article presents a PRISMA 2020-guided systematic review of peer-reviewed agentic AI research (January 2015 – December 2025), with deliberate exclusion of preprint repositories. Searches across five databases (Scopus, IEEE Xplore, ACM Digital Library, SpringerLink, Web of Science) retrieved 1,874 records. Following deduplication (n = 291) and dual-reviewer screening (Cohen's κ = 0.83), 147 peer-reviewed studies were retained after structured quality appraisal. Six architectural dimensions are synthesised: Planning and Reasoning (n = 43); Tool Use and Environment Interaction (n = 46); Memory Architectures (n = 31); Self-Improvement and Meta-Reasoning (n = 22); Evaluation Benchmarks (n = 26); and Safety and Alignment (n = 22). Critical analysis reveals four structural tensions: autonomy versus safety verification; benchmark performance versus deployment generalisation; memory paradigm fragmentation; and governance-architecture operationalisation deficit. The Translational Readiness Taxonomy (TRT) — a novel five-level deployment maturity framework introduced in this review — is applied across eight core architectural components; no component achieves TRL 5, and self-improvement mechanisms attain only TRL 1–2. Tool use and planning dominate the corpus (60.6%), while safety contributions (15.0%) and evaluation frameworks (17.7%) remain critically underrepresented. A five-priority research roadmap with measurable success criteria and minimum reporting standards is proposed.