Autonomous and Agentic AI with Digital Twins for Resilient Transportation and Smart Logistics: A Systematic Review, Multi-Axis Taxonomy, and Evidence-Informed Human-in-the-Loop Reference Architecture

Transportation and logistics systems are increasingly moving beyond predictive artificial intelligence toward systems capable of selecting, coordinating, and initiating operational decisions. The objective of this review is to systematically characterize autonomous and agentic AI in transportation and smart logistics, examine its integration with digital twins and trustworthiness mechanisms, and synthesize a multi-axis taxonomy and an evidence-informed human-in-the-loop reference architecture. Following PRISMA 2020, we synthesized 27 full-text-reviewed primary studies published from January 2021 through July 2026. For mutually exclusive descriptive reporting under the predefined precedence rule, the corpus is represented by 11 single-agent autonomous systems (40.7%), six LLM-agentic systems (22.2%), five MARL systems (18.5%), and five twin-centred autonomous systems (18.5%). These reporting strata are used only to prevent double counting and do not replace the multi-axis characterization of agency profile, decision topology, reasoning mechanism, and digital-twin relation. Reinforcement learning or MARL is used in 15 studies (55.6%), while digital twins are used in six (22.2%). Despite substantial experimentation with autonomous decision-making, only three (11.1%) studies reach operational validation. Cybersecurity is explicitly addressed in two studies (7.4%), explainability in six (22.2%), and explicit human approval, override, or intervention in three (11.1%). None of the LLM-core primary studies combines language-agent reasoning with a synchronized operational digital twin, revealing an important corpus-level architectural gap. Methodological rigor, validation strength, and trustworthiness are assessed as distinct dimensions rather than combined into a composite quality score. Based on the evidence, the review develops a ten-dimension taxonomy, proposes an explicitly unvalidated seven-layer human-in-the-loop reference architecture with autonomy gating, and presents an author-developed sixteen-threat mapping informed mainly by supporting cybersecurity literature. Overall, the evidence indicates rapid diversification of autonomous and agentic systems but persistent gaps in operational validation, integrated assurance, cybersecurity, and empirically tested human oversight.