A Large Language Model-based Agent System for Interactive Large-Scale Project Scheduling

Abstract Advanced optimization techniques generate highly complex schedules for large-scale engineering projects. While large language models have recently emerged as a transformative alternative for project scheduling, two primary hurdles must be overcome to realize a high-precision framework: performance degradation in large-scale projects and the functional separation between planning and analysis. To address these challenges, we propose an on-premises autonomous large language model agent system designed to raise open-source models to a level of accuracy that makes them a practical choice for industrial scheduling. Orchestrated within a LangGraph-based framework that uses the reasoning and acting paradigm, the agent facilitates a virtuous cycle in which plan generation and analytical feedback reinforce each other. To handle large-scale projects accurately, the agent uses a graph retrieval-augmented generation approach that combines local and global search over a knowledge graph of the schedule. It pairs this retrieval with analysis, modification, and optimization tools. Retrieval extracts only the relevant portion of the schedule, while the tools run schedule-wide computations that retrieval alone cannot perform, such as aggregation across all activities in the schedule. In a case study on real-world schedule data, the framework raised the accuracy of open-source backbones to about 90% on the large-scale schedule. By contrast, tool-free baselines struggled regardless of model: open-source models could not fit the schedule within their context window, and even a long-context commercial model (Gemini) reached only 41%. This points to the absence of tools, rather than the context window’s capacity, as the main bottleneck, a finding our component-wise ablation supports. Although this work was evaluated on a single project type at one shipyard, it demonstrates how large language models can become intelligent partners. It also provides a reference point for artificial intelligence-driven support in industrial scheduling and for similar work in other domains.