Abstract
<jats:p>The optimization of reaction conditions to max-imize product yield is a critical bottleneck in chemical innovation, where every additional iteration, in the lab or the manufacturing plant, adds to costs of reagents, time, and expert effort. We report the deployment of a Vision Language Model (VLM) agent in a real wet-chemistry laboratory for the iterative optimiza-tion of nickel-catalyzed C-glycosylation, operating under a tight, fixed budget. To ensure that every recommendation is executable, the agent selects conditions from an expert-curated candidate space in which each reagent is verified to be commercially procurable. Unlike Bayesian optimization, which relies on numer-ical featurization, lacks chemical priors, and produces black-box scores, our agent was made to natively process reaction-scheme images and reason in natural language. The improved align-ment with the way chemistry knowledge is actually expressed in the literature improves the recommendations. To elicit auditable recom-mendations, we introduce a sequential controlvariable design that optimizes one condition at a time, improving its trustworthiness for realworld laboratory deployment. In deployment, the system, compared with baseline methods, reached the highest yield faster under the same budget. We report deployment lessons on how the visual modality lowers the adoption barrier for chemists, the strengths and limitations of VLM reasoning revealed by expert evaluation, and the human-VLM collaboration patterns that emerged in practice.</jats:p>