We study peer effects in crime by analyzing co-offending networks. We first provide a credible estimate of peer effects in these networks equal to 0.17. This estimate implies a social multiplier of 1.2 for those individuals linked to only one co-offender and a social multiplier of 2 for those linked to three co-offenders. We then provide one of the first empirical tests of the key player policy in a real world setting. This policy defines a micro-founded strategy for removing the criminal from each network that reduces total crime by the largest amount. Using longitudinal data, we are able to compare the theoretical predictions of the key player policy with real world outcomes. By focusing on networks for which the key player has disappeared over time, we show that the theoretical predicted crime reduction is close to what is observed in the real world. We also show that the key player policy outperforms other reasonable police policies such as targeting the most active criminals or targeting criminals who have the highest betweenness or eigenvector centrality in the network. This indicates that behavioral-based policies can be more efficient in reducing crime than those based on algorithms that have no micro-foundation.