Agentic Reasoning Is Not Always Better: Evidence from Clinical Time-Series Explanation

Abstract

Continuous physiological monitoring is central to emergency care, yet building trustworthy, clinically usable AI remains challenging. While LLMs translate complex physiological signals into natural-language clinical narratives, it remains unclear whether agentic reasoning frameworks improve the quality of explanations over zero-shot inference. We present Vivaldi, a role-structured multi-agent system for explaining multivariate physiological time series in emergency department settings. We evaluate through a controlled clinical pilot involving a small cohort of emergency medicine experts. Their assessments challenge the assumption that increasingly complex agentic reasoning uniformly improves performance. Our experiments show agentic pipelines substantially benefit non-reasoning and medically fine-tuned models, improving expert-rated explanation justification and relevance by +6.9 and +9.7 points, respectively. Contrarily, for reasoning models, agentic orchestration often degrades explanation quality, including a 14-point drop in relevance, while improving diagnostic precision (ESI F1 +3.6). We also find explicit tool-based computation is decisive for codifiable clinical metrics, whereas subjective targets, such as pain scores and length of stay, show limited or inconsistent changes. Expert evaluation indicates gains in clinical utility depend on visualization conventions, with medically specialized models achieving favorable trade-offs between utility and clarity. Together, these findings show that the value of agentic AI lies in the selective externalization of computation and structure rather than in maximal reasoning complexity, highlighting concrete design trade-offs that are broadly applicable to explainable AI in safety-critical healthcare settings.