This study introduces a data-driven agent-based evacuation framework implemented in a building digital-twin environment to analyze the performance of static and dynamic evacuation guidance strategies under varying congestion conditions. The first three floors of Building H at the Osaka Ibaraki Campus (OIC) of Ritsumeikan University are used as a real-world testbed and are reconstructed, using Blender and Unity. Human movements are simulated via NavMesh-based navigation, incorporating empirically grounded heterogeneity in reaction times and walking speeds, together with a lightweight mechanism that adaptively reduced speed, based on local density. Seven guidance strategies are evaluated under measured off-peak and peak occupancy conditions. Results show that strategy effectiveness depends strongly on the congestion level. In off-peak scenarios, the Dynamic-Individual strategy achieved the shortest evacuation time, demonstrating the benefit of individual-level rerouting when viable alternative paths exist. Under peak conditions, however, dynamic strategies deteriorated markedly, particularly Dynamic-Individual, and evacuation times increased by approximately 72%, as saturated exits and staircases constrained rerouting opportunities. In contrast, pre-balanced static strategies, particularly Static-Phased, exhibited greater robustness under heavy congestion. While Static-Naive enables rapid early evacuation, Static-Phased reduced extreme tail delays and shortened the maximum evacuation completion time by approximately 10% compared to baseline, Static-Naive. These findings clarify when the dynamic guidance is effective and when the static guidance is preferable, and they underscore the broader value of digital-twin-based simulation as a generalizable tool for evacuation analysis and building safety planning.