Abstract
Agentic Artificial Intelligence (Agentic AI) has emerged as a new paradigm that extends conventional artificial intelligence by integrating autonomous reasoning, planning, memory, tool use, and adaptive decision-making into goal-directed systems. Although interest in Agentic AI has rapidly increased, its application within the construction industry remains fragmented and lacks a comprehensive synthesis. This study presents a systematic literature review of Agentic AI in construction, following the PRISMA methodology and analysing 50 journal articles published between 2021 and 2026. The review examines the conceptual evolution of Agentic AI, architectural paradigms, intelligence mechanisms, enabling technologies, lifecycle applications, reported benefits, validation approaches, adoption barriers, and future research directions. The findings show a clear transition from precursor forms of computational agency, including rule-based agent-based modelling and reinforcement-learning systems, towards contemporary Agentic AI characterised by increasingly modular, hybrid, tool-enabled, multi-agent, and human-governed architectures supported by large language models, digital twins, robotics, and Internet of Things technologies. Current applications primarily support project planning, scheduling, logistics, safety management, inspection, quality control, structural design, and facility operation. However, most studies remain at prototype or simulation stages, with limited real-world deployment and inconsistent evaluation methods. The review also identifies significant technical, organisational, and governance challenges, including engineering reliability, interoperability, explainability, human oversight, and data integration. Building on these findings, the study proposes the Human-Governed Construction Agentic Intelligence Loop (H-CAIL), a construction-specific closed-loop framework that integrates project context, multimodal perception, goal-directed planning, reasoning, tool selection, human-governed action, evaluation, and adaptive learning to support trustworthy autonomous decision-making across the construction lifecycle. Finally, the study presents a future research agenda to support the development of trustworthy, scalable, and construction-specific Agentic AI ecosystems capable of enabling autonomous yet human-governed project delivery across the entire asset lifecycle.