Closed-loop experimentation replaces the human step in the formulation cycle with an algorithm that selects the next composition and hardware that prepares and measures it without further instruction. This review assesses what that idea has delivered in pharmaceutics, and separates the three capabilities a working loop requires: a decision policy that is sample-efficient under uncertainty, execution hardware that can act without a person in the path, and a data layer that makes every outcome, including every failure, machine-readable. Evidence is now solid for the first. Bayesian optimization reached optimal orally disintegrating tablet conditions in roughly ten experiments where a factorial design needed about twenty-five, produced a monoclonal antibody formulation optimized simultaneously for thermal stability, colloidal interaction and interfacial stress in thirty-three experiments, and located high-solubility injectable vehicles after sampling 256 of 7,776 combinations. A robotic tableting platform has carried six active ingredients from raw material characterisation to in-specification tablets within six hours while cutting material use by 65%. The second capability is advancing but remains limited by the physical awkwardness of pharmaceutical matter, particularly cohesive powders and low-dose solids. The third is largely absent. The central argument is that the binding constraint is metrological rather than computational: dissolution behaviour, physical stability over months and performance in a patient cannot be measured on the timescale a loop consumes, so every autonomous campaign optimizes a surrogate. Managing the distance between surrogate and endpoint, together with reporting sample efficiency against stated baselines, will decide whether these systems produce better medicines or only faster experiments.