How 10 Hospitals Used a Multi-Agent System on AWS Bedrock to Cut Discharge Delays by 35%

A 10-Hospital Multi-Agent Pilot for Patient Flow Optimization and Discharge Planning on AWS Bedrock As of May 24, 2026, a consortium of 10 major hospital systems completed the first known multi-agent pilot for patient flow optimization and discharge planning on AWS Bedrock. The pilot, which ran for 12 weeks across diverse clinical settings, demonstrated that pairing a clinical decision support agent (Qwen 3.8 Max) with a scheduling agent (Llama 5) can yield measurable operational improvements: 35% faster patient throughput and a 20% reduction in administrative overhead. This article provides a vendor-neutral examination of the architecture, metrics, and a replication blueprint for healthcare leaders evaluating multi-agent systems. What Does a 10-Hospital Multi-Agent Pilot Actually Look Like? A consortium of 10 U.S. hospital systems — ranging from academic medical centers to community hos

pitals — came together under a shared data governance agreement to test whether multi-agent AI could improve discharge planning. Each hospital had existing investments in electronic health records (EHR) and varying levels of digital maturity, but none had deployed multi-agent systems at scale. The pilot was orchestrated on AWS Bedrock, leveraging its native multi-agent collaboration and model hosting capabilities. The pilot focused on the final 48 hours of the inpatient stay: a phase identified as the most variable and prone to delays. Each hospital integrated two specialized agents into their existing workflows via secure API calls. The consortium did not share protected health information (PHI) between sites; each hospital ran the same agent architecture on its own tenant within AWS Bedrock. The results were aggregated from anonymized dashboards. The Architecture: Qwen 3.8 Max for Clin

ical Decisions, Llama 5 for Scheduling The multi-agent system on AWS Bedrock used two purpose-built models orchestrated by Bedrock Agents: Qwen 3.8 Max (from the Qwen model family on Hugging Face) handled clinical decision support. It analyzed structured EHR data — lab results, medications, comorbidities, and care plan updates — to generate a daily readiness-to-discharge score and flag potential barriers (e.g., pending consults, unstable vitals, need for home health). Qwen 3.8 Max runs on a Mixture-of-Experts architecture with 3.8 trillion parameters optimized for reasoning and domain-specific tasks. In this pilot, it was fine-tuned on de-identified clinical notes to improve adherence to discharge criteria. Llama 5 (from Meta) managed scheduling and logistics. Given a list of patients flagged as discharge-ready by Qwen 3.8 Max, Llama 5 coordinated tasks: ordering transportation, notifyin

g family, scheduling follow-up appointments, and aligning pharmacy pickup windows. Llama 5’s strong instruction-following and planning capabilities allowed it to manage up to 200 parallel schedules per hospital. The two agents communicated through a shared state store in Amazon DynamoDB, with Bedrock Agents handling routing and error recovery. A human-in-the-loop approval step was required for all clinical decisions (Qwen recommendations were presented to the attending physician), while scheduling actions from Llama 5 executed automatically unless overridden. Metrics That Matter: 35% Faster Throughput and 20% Admin Reduction The consortium tracked five operational metrics before and after the pilot. The key results, averaged across all 10 hospitals, are associated with the deployment of the multi-agent system: Metric Baseline (pre-pilot) Pilot period Improvement :------------------------

----------------------------- :------------------- :----------- :--------------- Average time from discharge order to patient departure 4.2 hours 2.7 hours 35% faster Administrative hours spent per discharge (data entry, calls, paperwork) 1.5 hours 1.2 hours 20% reduction Rate of discharge-related readmissions within 7 days 4.8% 4.5% Not significant Physician satisfaction with discharge workflow (1-10 scale) 6.2 7.8 +1.6 improvement Patient satisfaction with discharge education (1-10 scale) 7.0 8.1 +1.1 improvement The throughput improvement was driven by earlier identification of discharge-ready patients (Qwen 3.8 Max flagged candidates an average of 2.1 hours sooner than manual review) and parallel task execution by Llama 5. The administrative reduction came from automating routine calls and documentation – tasks that previously required nurses to manually coordinate. How the Consortiu

m Built the Multi-Agent System in 12 Weeks Development followed an agile, consortium-wide sprint schedule: 1. Week 1-2: Governance and data mapping. Establish shared data definitions, PHI isolation boundaries, and model access policies. Each hospital mapped its EHR fields to a common schema. 2. Week