How a 15-Hospital Consortium Cut Length of Stay by 12% with Multi-Agent AI on AWS Bedrock

The Multi-Agent Pilot at a 15-Hospital Consortium: Background and Goals The consortium—a group of 15 mid-sized hospitals in the United States—launched the pilot in Q3 2025 to address two persistent operational pain points: unpredictable patient discharge timelines and inefficient inventory management of surgical supplies. Traditional forecasting models using linear regression for patient flow had a MAPE (mean absolute percentage error) of 28%, leading to overstocking of high-cost supplies and understaffing of beds. The goal was to reduce these inefficiencies by deploying a multi-agent system that could forecast patient flow in real-time, dynamically allocate resources (beds, staff, supplies), and coordinate decisions across departments. The consortium chose AWS Bedrock for its managed agent collaboration capabilities and long-term security compliance for healthcare data. The pilot focuse

d on surgical units—a high-cost, high-variability area—and ran for 8 months with full HIPAA compliance. Architecture Overview: Patient Flow, Resource Allocation, and Coordination Agents The multi-agent system comprised three distinct agents orchestrated via AWS Bedrock AgentCore: - Patient Flow Forecasting Agent (Qwen 3.8 Max): This agent ingested real-time data from EHR, including admission times, diagnosis codes, procedure schedules, nursing shift notes, and historical discharge patterns. Using Qwen 3.8 Max's time-series capabilities, it generated probabilistic forecasts of bed demand and expected length of stay per patient, updated every 2 hours. - Resource Allocation Optimization Agent (Llama 5): Llama 5 received the forecasts and current inventory data from the hospital's supply chain management system. It solved a multi-constraint optimization problem to assign beds, operating room

slots, and equipment while minimizing waste of high-cost consumables (e.g., surgical kits, implantables). Llama 5 was chosen for its strong reasoning over structured decisions. - Coordination Agent (custom, based on Amazon Bedrock orchestration): This agent mediated between the first two. It aligned the patient flow forecast with resource availability, resolved conflicts (e.g., a predicted surge conflicting with low PPE stock), and triggered alerts to department heads. It also logged decisions for audit trails. The agents communicated through Bedrock's multi-agent collaboration, which handled message routing and state management. Each agent had a dedicated knowledge base of hospital policies and historical constraints, stored in encrypted vector databases on AWS. Model Selection: Why Qwen 3.8 Max and Llama 5 Were Chosen The consortium evaluated several models for each agent role: - Qwen

3.8 Max (Alibaba Cloud): Selected for patient flow forecasting because its architecture handles long-context time-series data (up to 128K tokens) and excels at trend extrapolation. In benchmarks from the Qwen 3.8 Max model card (May 2025), it achieved state-of-the-art results on the MSTL dataset for multivariate time series. Its cost for inference on AWS Bedrock was $0.15 per 1M input tokens (as of May 2026), well within the consortium's budget. - Llama 5 (Meta): Chosen for resource allocation optimization because of its strong performance on integer linear programming-style tasks encoded in natural language. The Llama 5 release page (April 2026) highlighted a 15% improvement over Llama 4 on the MATH and GSM8K benchmarks, critical for reasoning about constraints like bed capacity per unit, nurse-to-patient ratios, and shelf-life of supplies. Llama 5's 70B variant ran efficiently on Bedr

ock's provisioned throughput. - Both models are open-weight, allowing the consortium to customize with LoRA fine-tuning on hospital-specific de-identified data. This flexibility was a major factor, as proprietary alternatives would have required full data-sharing with closed vendors. The consortium did not use a single massive model due to latency and cost—splitting tasks across specialized agents reduced average response time by 40% compared to a monolithic GPT-4-level system. Integrating with Existing EHR Systems: Data Pipelines and Security EHR integration was the most complex part of the deployment. The hospitals used a mix of Epic and Cerner systems. The data pipeline followed a HIPAA-compliant architecture: - Data extraction: A one-way push from EHR's FHIR API into an Amazon S3 bucket in the same AWS region, encrypted at rest with KMS. Only de-identified fields (no PHI in names or

social security numbers) were used: timestamps, department codes, bed occupancy, supply usage logs. - Data transformation: AWS Glue ETL jobs cleaned and aggregated the data into hourly features for the agents. The patient flow agent received rolling 7-day windows of admission, discharge, and transfe