We develop a model of repeated multilateral bargaining that links cycles via the identity of the agenda setter. In sharp contrast to the standard history-independent equilibrium predictions, in an experiment, we observe stable and persistent coalitions in terms of member identity, allocations to coalition partners, and agenda-setter identity. Our results call into question the validity of restricting attention to static, history-independent strategies in dynamic bargaining games. We show that weakening the standard equilibria concepts to allow players to condition on a single piece of history is enough to generate equilibria which are consistent with observed laboratory outcomes.