A number of theoretical results have provided sufficient conditions for the selection of payoff-efficient equilibria in games played on networks when agents imitate successful neighbors and make occasional mistakes (stochastic stability). However, those results only guarantee full convergence in the long-run, which might be too restrictive in reality. Here, we employ a more gradual approach relying on agent-based simulations avoiding the double limit underlying these analytical results. We focus on the circular-city model, for which a sufficient condition on the population size relative to the neighborhood size was identified by Alós-Ferrer & Weidenholzer [(2006) Economics Letters, 93, 163–168]. Using more than 100,000 agent-based simulations, we find that selection of the efficient equilibrium prevails also for a large set of parameters violating the previously identified condition. Interestingly, the extent to which efficiency obtains decreases gradually as one moves away from the boundary of this condition.