Abstract
<jats:p>Objectives. Construction management (CM) tools are increasingly marketed as “agentic”, but the term covers systems with very different levels of autonomy and reliability, leaving CM organizations without a shared vocabulary for evaluating what a tool does or where it may fail. This paper translates a general-purpose taxonomy of agentic artificial intelligence (AI) design patterns into construction-specific vocabulary so practitioners can reason about pattern selection deliberately rather than adopting agentic tooling as a black box. Methods. Twenty patterns were extracted from a practitioner reference on agentic system architecture and mapped onto recurring CM information workflows, each paired with a brief illustrative case grounded in transportation construction. One pattern, Retrieval, was validated empirically using a synthetic bridge rehabilitation specification corpus (16 base sections, 6 addenda) and a 24-question test set (stable, addendum-affected, and distractor types), evaluated across three indexing conditions and two chunking strategies. Findings. The twenty patterns were organized into four functional groups, each illustrated with an original architecture diagram. The retrieval case study found that an index built only from original specifications answered 0 percent of addendum-affected questions correctly; adding current addenda raised this to 50 percent, and supersession-aware re-ranking raised it to 100 percent but reduced distractor accuracy from 87.5 to 62.5 percent, showing that grounding requires more than re-indexing alone. Novelty. This is the first systematic translation of the general-purpose agentic AI design pattern literature into CM vocabulary, and the first empirical test of the Retrieval pattern's addenda-grounding claim against a working pipeline. Practical Applications. The taxonomy gives CM digital transformation leads and procurement teams a common language for evaluating vendor claims, while the case study offers a reproducible method for auditing whether a retrieval tool is actually grounded in current project documents.</jats:p>