Build closed loops where agents train detectors and classifiers, test validation sets, and improve models with little human input.