EnzymeRSI implements a dual-loop recursive self-improvement (RSI) paradigm. The outer loop evolves “harness assets” without touching model weights, while the inner loop refines prediction and selection algorithms under controlled settings. A cloud-lab module closes the physical loop by feeding quality-controlled experimental feedback. By iteratively comparing predictions, the outer loop raised the mean Spearman correlation on ProteinGym by 0.044 across five assays, reaching 0.335 on CD19 and 0.683 on OPSD (rhodopsin); CD19’s top rank was produced entirely by the outer loop. The inner loop self-improved ThermoMPNN to a Pearson correlation of 0.849 and achieved a Spearman of 0.307 on TadA-Bench within 1.13 h. Three prospective wet-lab tests demonstrated the system’s decision-making: among 32 unscheduled luciferase (RLuc) candidates, 10 showed ≥2× wild-type activity; a suggested A3A cytidine deaminase variant raised 5mC conversion to 90.2%; and path selection in evolution increased the catalytic efficiency of DNA polymerase KOD-E on a modified substrate to 31.1-fold of the parent. By treating workflow organization and algorithm development as a single falsifiable dry-to-wet loop, this work offers an evolutionary paradigm for moving beyond the blind spots of fixed methods toward autonomous enzyme engineering. Review