Beyond Model Choice: A Harness-Centered Framework for Designing Clinical Large Language Model Agents

Clinical large language model agents depend on both model capability and the surrounding system that selects information, constrains actions, and verifies execution. In this Viewpoint, we argue that this harness should be an explicit unit of clinical agent design, evaluation, and reporting. Our aim is to help clinical informatics and development teams specify this layer and to help evaluators and institutional governance teams assess its contributions and risks. Selected benchmarks show substantial performance differences across harness configurations, but architecture comparisons, bundled improvements, and within-model component comparisons support different levels of causal attribution. We propose a revisable 3-tier framework: tier 1 specifies general orchestration functions, including prompts, tools, output schemas, control flow, memory, and generic verification; tier 2 specifies their clinical adaptations for knowledge integration, differential reasoning, missingness and temporality, safety verification, and workflow alignment; and tier 3 specifies a failure-driven evaluation cycle applied to both. Our contribution is a framework that makes these connections explicit. We provide classification rules, an illustrative clinical example, a proposed evaluation protocol, and a minimum reporting checklist. Observable execution traces are distinguished from private model reasoning, with provenance and privacy safeguards supporting retrospective reconstruction. Evaluation should separate task success, clinical correctness, process safety, and operational reliability. Lightweight fine-tuning can complement harness design; their interaction requires controlled study. The framework is a proposal for empirical testing, not a validated taxonomy or evidence of deployment safety.