CLOUD-NATIVE ORCHESTRATION PATTERNS FOR MULTI-AGENT HEALTHCARE LLMS

Abstract

Healthcare organizations increasingly deploy multi-agent large language models (LLMs) for clinical triage, diagnostic support, and documentation. However, deployment at scale requires coordination across distributed cloud infrastructure while satisfying HIPAA, GDPR, and HL7 FHIR interoperability requirements. We evaluated three orchestration patterns—Kubernetes-native microservices, serverless Knative, and hybrid edge-cloud—for multi-agent healthcare LLM orchestration using synthetic clinical data (10,000 patients, 45,000 encounters via Synthea). Multi-cloud experiments across AWS, Azure, and GCP (15 repetitions per pattern; n=405 total) demonstrate that edge-local deployment reduces latency 45% (triage: 2.08s vs. 3.81s) and maintains 99.97% compliance adherence (±0.03%) across 1.2 million policy-controlled PHI accesses. Multi-agent diagnostic consensus improved accuracy from 82.8% (single-agent) to 87.1% (ANOVA: F(2,403)=18.4, p<0.001, η²=0.20) with greatest benefits for complex multi-system conditions. Clinical documentation ratings by three board-certified physicians (κ=0.72 inter-rater agreement) achieved 4.15/5 mean acceptability on synthetic cases, though generalization to real clinical practice requires prospective validation. We propose a reference architecture integrating Istio service mesh (mTLS, policy enforcement), Knative event-driven activation, and Kubernetes-native microservices, establishing design guidelines for healthcare organizations selecting orchestration patterns based on latency, cost, and compliance constraints. Limitations include reliance on synthetic ground truth, single-timepoint evaluation, and unknown generalization to alternative LLM architectures.