The typically large numbers of interactions in agent-based simulations come at considerable computational costs. In this article, we present an approach to reduce the number of interactions based on behavioural patterns that recur during runtime. We employ machine learning techniques to abstract the behaviour of groups of agents to cut down computational complexity while preserving the inherent flexibility of agent-based models. The learned abstractions, which subsume the underlying model agents’ interactions, are constantly tested for their validity: after all, the dynamics of a system may change over time to such an extent that previously learned patterns would not reoccur. An invalid abstraction is, therefore, removed again from the system. The creation and removal of abstractions continues throughout the course of a simulation in order to ensure an adequate adaptation to the system dynamics. Experimental results on biological agent-based simulations show that our proposed approach can successfully reduce the computational complexity during the simulation while maintaining the freedom of arbitrary interactions.