This paper reports the results of two randomized field experiments, each offering different populations of youth a supported summer job in Chicago. In both experiments, the program dramatically reduces violent-crime arrests, even after the summer. It does so without improving employment, schooling, or other types of crime; if anything, property crime increases over 2-3 post-program years. To explore mechanisms, we implement a machine learning method that predicts treatment heterogeneity using observables. The method identifies a subgroup of youth with positive employment impacts, whose characteristics differ from the disconnected youth served in most employment programs. We find that employment benefiters commit more property crime than their control counterparts, and non-benefiters also show a decline in violent crime. These results do not seem consistent with typical theory about improved human capital and better labor market opportunities creating a higher opportunity cost of crime, or even with the idea that these programs just keep youth busy. We discuss several alternative mechanisms, concluding that brief youth employment programs can generate substantively important behavioral change, but for different outcomes, different youth, and different reasons than those most often considered in the literature.