JMIR Publications

ABSTRACT

Large language model (LLM) agents are increasingly deployed in clinical settings to execute multistep tasks such as electronic health record retrieval, clinical decision support, and patient communication. Discussion of these systems has focused largely on which foundation model performs best on medical benchmarks. This focus understates the role of the harness, the engineered scaffolding around the model that defines system prompts, tool interfaces, control flow, output schemas, memory, and verification logic. Recent benchmarks demonstrate that for medical agent tasks, harness design can change task success rates by tens of percentage points while the underlying model is held constant, and that the choice of agent architecture often outweighs the choice of base model. Although recent work has surveyed medical agent architectures at the system level, proposed agent engineering frameworks for biomedical research, and articulated governance competencies for clinicians supervising agents, the harness layer between these levels has received limited systematic attention. We propose a three-tier harness-centered framework: tier 1 specifies foundational harness components shared with general-domain agents (system prompts, tool definitions, output schemas, control flow, memory); tier 2 introduces medical-specific harness components addressing clinical knowledge integration, differential reasoning structure, missing-data handling, safety verification, and clinical workflow alignment; tier 3 establishes failure-mode-driven adaptation through systematic trace analysis, taxonomy-guided redesign, and cross-benchmark validation. We illustrate each tier with evidence from recent medical agent benchmarks, discuss operationalization for development teams, address an alternative view that frames harness engineering as a temporary patch superseded by future fine-tuning, and identify open questions for the field.