Scope, Challenges, and Future of Agentic AI in Geotechnical and Earthquake Engineering: A Critical Review

Agentic artificial intelligence (agentic AI) refers to systems that pursue goals by planning, invoking external tools, maintaining memory and acting on the results with a defined degree of autonomy. This review examines the scope, challenges and future directions of agentic AI in geotechnical and earthquake engineering, with an explicit distinction between capabilities that have been demonstrated and those that remain conceptual. We first propose a taxonomy that separates predictive machine learning, generative AI, retrieval-augmented language-model systems, tool-using agents, multi-agent systems and closed-loop autonomous agents, and we distinguish within-episode feedback and memory updating from online learning, reinforcement learning and model retraining. We then synthesize 22 studies (2023–2026) in which large language models (LLMs) or agents perform geotechnical, earthquake or adjacent structural engineering tasks, together with two closed-loop exemplars from chemistry, and classify each by architecture, tools, memory, validation, uncertainty treatment, human involvement and evidence level. All civil-engineering studies located are prototypes or proof-of-concept systems; no field-validated agentic deployment in geotechnical or earthquake practice was found, and no peer-reviewed study of an agent operating a commercial geotechnical finite-element package was identified. We discuss the enabling layers that engineering agents require—retrieval-augmented generation over codes and site records, knowledge graphs and data standards (AGS4, DIGGS, IFC), interoperability with OpenSees, PLAXIS and OpenQuake, and physics-informed modeling—and report indicative cost and latency figures. An illustrative proof-of-concept of the tool layer for CPT-based liquefaction assessment shows how deterministic triggering analysis, Monte Carlo uncertainty propagation and human approval gates can be combined. Finally, we assess eight challenges with an explicit ordinal rubric and present a conditional short-, medium- and long-term roadmap linked to eight research gaps. Agentic AI can add value as a supervised orchestration layer over validated engineering tools; claims of autonomous engineering practice are not yet supported by evidence.