Agentic supervision of iterative assay design in high-throughput regulatory genomics

Designing maximally informative experiments remains a central challenge in genomics. In regulatory genomics, massively parallel reporter assays (MPRAs) enable large-scale functional measurements but require carefully constructed sequence libraries to be maximally informative. We introduce an agent-based framework for MPRA library design in which a large language model autonomously constructs and improves MPRA libraries designed with the goal of training predictive models of regulatory activity. Candidate libraries are evaluated using a high-fidelity surrogate of MPRA measurements, enabling rapid closed-loop optimization under realistic experimental constraints in a simulated wet-lab/dry-lab loop. Across independent runs, agent-designed libraries outperform a set of human-designed strategies and yield consistent improvements in predictive performance. In addition, the agent recovers interpretable design principles. These results suggest that AI agents can contribute meaningfully to experimental design in regulatory genomics and provide a reusable \textit{in silico} benchmark for evaluating AI co-scientist workflows.