Agentic Systems for Sample-Efficient Drug Formulation Design

Self-emulsifying drug delivery systems (SEDDS) can enhance the oral bioavailability of poorly soluble drugs, but their development remains constrained by how formulation space is explored in practice. The problem is not only that the space is large, but that existing workflows typically explore only a small, familiar, and conservative subset of it, restricting innovation and limiting opportunities to discover differentiated formulations. Traditional approaches based on excipient screening, pseudo-ternary phase diagrams, and design-of-experiments (DoE) are labor-intensive and better suited to local refinement than to broad, discovery-oriented exploration. Bayesian optimization (BO) has improved experimental efficiency, but still depends on multiple sequential rounds of wet-lab testing. In this work, we present a closed-loop agentic system, grounded in structured experimental data, that autonomously designs SEDDS formulations and executes them in an integrated miniaturized laboratory where preparation, physico-chemical characterization and dispersion assay run end-to-end, evaluating 128 formulations in just over a week of platform time. The system matches BO performance with 4$\times$ fewer experiments. An ablation removing the experimental data reduces optimization performance by 44\%, establishing that structured experimental evidence, rather than LLM parametric knowledge alone, drives the advantage. A second agentic batch improves formulation quality by 33\% over the first batch's results, with no model retraining. Taken together, these results are a first step toward practical and scalable autonomous formulation design, enabling broader exploration of formulation space with substantially reduced experimental burden.