We show that an agentic large language model (LLM) (OpenAI o3 with deep research) can autonomously reason, write code, and iteratively refine hypotheses to derive a physically interpretable equation for competitive adsorption on metal-organic layers (MOLs)—an open problem our lab had struggled with for months. In a single 29-min session, o3 formulated the governing equations, generated fitting scripts, diagnosed shortcomings, and produced a compact three-parameter model that quantitatively matches experiments across a dozen carboxylic acids.