LLM-Based Multiagent System for Construction Job Safety Analysis

Job safety analysis (JSA) is a safety management framework for the construction sector that relies on manual processes and is constrained by the limitations of manual performance. Although single-agent large language model (LLM) applications have shown potential for automating aspects of safety management, they lack contextual depth and multistep reasoning. This study introduces an LLM-based multiagent system to automate JSA and enhance construction safety management. This system employs multiple agents, each specialized in distinct phases of JSA. The systematic design, development, and evaluation of the framework highlight the advantages of the multiagent approach. Compared with alternative approaches, the system produced more accurate and comprehensive safety risk identification, assessment, and management. Both automated evaluation and expert reviewers confirmed its effectiveness and practical applicability in advancing safety management practices. Furthermore, several key parameters, such as prompt engineering, fine-tuning, and the integration of feedback loops, and their influence on the performance of multiagent systems are explored. This study offers a range of applicable insights for designing high-performance multiagent systems in the construction safety domain, considering factors such as task complexity, cost, and time.