Agent-based models of political party competition in a multidimensional policy space have been developed in order to reflect adaptive learning by party leaders with very limited information feedback. The key assumption is that two categories of actors continually make decisions: voters choose which party to support and party leaders offer citizens a certain policy package. After reviewing the arguments for using agent-based models, I elaborate two ways forward in the development of these models for political party competition. Firstly, theoretical progress is made in this article by taking the role of the mass media into account. In previous work it is implicitly assumed that all parties are equally visible for citizens, whereas I will start from the more realistic assumption that there is also competition for attention in the public sphere. With this addition, it is possible to address the question why new parties are seldom able to successfully compete with political actors already within the political system. Secondly, I argue that, if we really want to learn useful lessons from simulations, we should seek to empirically falsify models by confronting outcomes with real data. So far, most of the agent-based models of party competition have been an exclusively theoretical exercise. Therefore, I evaluate the empirical relevance of different simulations of Dutch party competition in the period from May 1998 until May 2002. Using independent data on party positions, I measure the extent to which simulations generate mean party sizes that resemble public opinion polls. The results demonstrate that it is feasible and realistic to simulate party competition in the Netherlands with agent-based models, even when a rather unstable period is investigated.