This paper analyzes the formation of partnerships in social networks. Agents randomly request favors and turn to their neighbors to form a partnership. If favors are costly, agents have an incentive to delay the formation of the partnership. In that case, for any initial social network, the unique Markov Perfect equilibrium results in the formation of the maximum number of partnerships when players become infinitely patient. If favors provide benefits, agents rush to form partnerships at the cost of disconnecting other agents and the only perfect initial networks for which the maximum number of partnerships are formed are the complete and complete bipartite networks. The theoretical model is tested in the lab. Subjects generally play according to their equilibrium strategy and the efficient outcome is obtained over 78% of the times. Decisions are affected by the complexity of the network. Two behavioral rules are observed during the experiment: subjects accept the formation of the partnership too often and reject partnership offers when one of their neighbors is only connected to them. JEL Classification Numbers: D85, C78, C91 Keywords: social networks, partnerships, matchings in networks, non-stationary networks, laboratory experiments * We thank Yann Bramoullé, Gabrielle Demange, Matt Eliott, Sanjeev Goyal, Matt Jackson, Eva Tardos and seminar participants at different institutions for their comments on the paper. We are grateful to Quentin Thevenet for programming the experiment and to Yannick Boireaud for valuable research assistance.