MedOps-RiskAdapt: Risk-Adaptive Cloud-Native Orchestration for Multi-Agent Healthcare LLMs under Workload Drift

Abstract

Multi-agent large language model (LLM) systems are increasingly considered for healthcare documentation, triage support, policy-aware reasoning, and operational coordination. Their deployment differs from general LLM serving because clinical workloads are non-stationary, patient-facing latency is safety relevant, and every inter-agent data exchange must satisfy privacy, audit, and interoperability constraints. This paper presents MedOps-RiskAdapt, a risk-adaptive cloud-native orchestration framework that combines workload-drift sensing, policy-gated agent routing, and adaptive cloud/edge placement for healthcare LLM agents. Unlike static Kubernetes or reactive autoscaling baselines, MedOps-RiskAdapt changes routing and escalation policy according to encounter risk, workload surge, and audit completeness signals. We evaluate the framework in a reproducible controlled benchmark using a Synthea-style synthetic patient cohort, three workload classes, two regulatory regions, three drift regimes, and 15 random seeds, producing 810 orchestration runs and 64,800 encounter-equivalents. Results show that MedOps-RiskAdapt reduces mean median latency from 6.42 s under static Kubernetes to 4.89 s, improves throughput from 166.7 to 213.1 encounters/min, and increases clinical task F1 from 80.7% to 84.4%. The p95 latency remains sensitive to surge conditions, rising to 12.55 s under MedOps-RiskAdapt, which is lower than both static Kubernetes and reactive HPA but not eliminated. Policy block rate improves from 98.2% to 99.3%, and audit completeness improves from 97.1% to 98.7%. The findings indicate that risk-adaptive orchestration can improve healthcare LLM infrastructure efficiency and governance, while residual tail latency and synthetic-data limitations remain important deployment constraints.