Dynamic systems, where the behaviour is the dominant characteristic, pose engineering challenges that are often neglected in model-based software engineering. However, supporting simulation development from design models is important in demonstrating that a simulator and simulation experiments are fit for their intended purpose. In engineering agent-based simulations, observable system behaviour is built up from the behaviour of low-level components; such simulations are used as research tools in (for instance) biological systems research. We have found that domain experts can validate diagrammatic models of behaviour and accompanying text, but we need model-based software development, and ideally automatable model transformation, to maintain fitness for purpose into code and experimentation. We present an exploration of behavioural model transformation, devising and applying manual transformation guidelines to an existing, published Java Mason simulator, created using the CoSMoS approach and UML-style state diagrams. We succeed in recreating part of the class structure of the code, but also expose many issues to be overcome, in terms of what needs to be modelled to enable transformation, and how and when design decisions are taken and documented. We also identify the need to generate the creation of low-level simulation, visualisation, and data capture mechanisms, and a means to design and encode simulation experimentation.