Learning to Evolve Clinical Harnesses Through Entity-Graph-Guided Orchestration for Clinical Decision Making

Large language model based agents show promise for clinical decision making but struggle to maintain coherent patient states and coordinate execution across long-horizon workflows. Harness optimization can improve this execution infrastructure without updating model parameters, yet loosely structured clinical evidence and fixed proposal strategies limit its effectiveness. Thus, we propose a framework that learns to evolve clinical harnesses by coupling entity-graph-guided harness with a self-evolving proposer. To represent evolving clinical evidence coherently and explicitly, a shared patient entity graph explicitly represents evolving clinical evidence, diagnostic hypotheses, and their dependencies, coordinating memory retrieval, clinical-skill selection, and tool execution through a consistent, stage-visible patient state. To mitigate optimization collapse arising from a fixed proposer, we introduce a self-evolving proposer that learns how to evolve the harness via Candidate-DAG memory and an evolving proposer skill library. Candidate-DAG memory preserves alternative harness designs, component inheritance, and case-level corrections and regressions, including useful interventions from lower-scoring candidates. An evolving proposer skill library distills this experience into reusable procedures for parent selection, cross-branch component reuse, avoidance of unsuccessful interventions, and targeted acquisition of training-derived medical knowledge. Extensive experiments on MedChain and ClinicalBench datasets demonstrate that our framework outperforms general-purpose and medical-specialized LLMs, multi-agent systems, and existing harness optimization methods in average performance.