Autonomous Agents for Auditable Cardiovascular Artificial Intelligence Development

Clinical artificial intelligence (AI) models are usually reported as finished artifacts, but each model reflects a limited human search across a much larger space of architectures, inputs, losses, optimizers, and training recipes. We tested whether autonomous code-writing agents could perform a controlled model-development experiment: proposing and evaluating code changes, and seeking performance gains without new data or human-guided edits. We built two such agents: an Iteration Agent that searches sequentially, keeping the best variant at each step, and an Evolution Agent that searches for variations in parallel using multiple large language models and prioritizes high-performing lineages across generations. In two architecturally distinct AI-enhanced electrocardiography (AI-ECG) models for structural heart disease, agent-optimized variants improved rank discrimination across held-out, external, and cross-institution evaluations, with area under the receiver operating characteristic curve gains of +0.006 to +0.039 (paired p < 0.05). At a fixed 90% sensitivity, specificity rose by up to 7.1 percentage points and positive predictive value by up to 4.8 percentage points. The selected code changes were substantive, spanning architecture, representation, and training recipe variations. These findings position autonomous agents as an auditable layer for clinical AI model improvement, provided that candidate selection, external validation, and post-update governance are explicit. We release these agents as an open, reusable toolkit.