In this paper, we are interested in ad hoc autonomous agent team composition using cooperative co-evolutionary algorithms (CCEA). In order to accurately capture the individual contribution of team agents, we propose to limit the number of agents which are updated in-between team evaluations. However, this raises two important problems with respect to (1) the cost of accurately estimating the marginal contribution of agents with respect to the team learning speed and (2) completing tasks where improving team performance requires multiple agents to update their policies in a synchronized manner. We introduce a CCEA algorithm that is capable of learning how to update just the right amount of agents' policies for the task at hand. We use a variation of the El Farol Bar problem, formulated as a multi-robot resource selection problem, to provide an experimental validation of the algorithms proposed.