AUTONOMOUS CLINICAL PATHWAYS: EVALUATING THE EFFICACY AND SAFETY OF AGENTIC AI ORCHESTRATION IN EMERGENCY DEPARTMENT TRIAGE

Emergency Department (ED) overcrowding is a persistent global healthcare challenge associated with increased morbidity, mortality, and clinician burnout. A major contributor to this crisis is the operational delay between patient presentation and clinical action—often described as the clinical action gap. While predictive artificial intelligence has improved risk detection, most systems lack the capacity to initiate clinical workflows, leaving triage processes largely manual and time intensive. This study evaluates the performance and safety of an agentic artificial intelligence orchestration framework designed to autonomously conduct patient intake interviews, assign Emergency Severity Index (ESI) levels, and initiate diagnostic orders through clinical information systems. A mixed-method evaluation was conducted consisting of a retrospective in-silico simulation involving 10,000 historical ED cases, and a prospective shadow clinical trial comparing AI-generated triage recommendations with real-world nurse triage decisions. The system used multimodal data fusion combining natural language patient interviews, real-time physiological vitals, and electronic health record data. The architecture employed retrieval-augmented generation (RAG) integrated with a multi-agent orchestration system capable of tool use through hospital APIs. Primary endpoints included inter-rater reliability (weighted Kappa) between AI and triage nurses, door-to-order latency, and safety performance for high-acuity cases. The agentic system demonstrated strong agreement with clinical triage, achieving a weighted κ = 0.89. Sensitivity for ESI Level 1 (resuscitation) cases reached 100%. Mean time from patient arrival to initial diagnostic order placement decreased from 28.4 minutes to 4.2 minutes (p < 0.001). Clinician feedback indicated reduced documentation burden but moderate workflow adaptation concerns. Agentic AI orchestration can significantly improve ED operational efficiency while maintaining clinical safety. By transitioning AI from predictive analytics to autonomous workflow orchestration, healthcare systems may reduce boarding times, improve patient throughput, and alleviate clinician cognitive load.