SAGE preprint reports fewer regressions when agents revise their workflows

A preprint introduces SAGE, a statistical gate for accepting changes to AI agents' persistent workflow instructions. Across five benchmarks and four underlying language models under equal budgets, the researchers report lower regression rates in 19 of 20 settings and an unchanged rate in the remaining setting. SAGE compares old and edited instructions on identical validation items, penalizes losses on previously solved items and accepts changes only when gains are statistically reliable. It achieved the highest final score in all 20 settings, raising LiveMath from 34.15 to 48.78. The demonstrated results concern benchmark evaluations, rather than deployed production agents.