When we focus on (Distributed) Artificial Intelligence as an experimental science, it is not always clear which are the best experiments to try and how they should be conducted. It is up to the designer of the experiment the choice from a wide range of languages, models and architectures for his agents, and even from a numerous set of computational environments (namely, workbenches and testbeds).
In the follow-up of some methodological remarks recently made, concerning the journey from the first idealization of a problem until a final system is completed and running, with this paper we aim to enlighten the importance of maintaining consistent and coherent links between levels of description. Only in this way, we can debug experiments at the higher levels of abstraction, and hope the implementation results we get are sound, in order to make the journey backwards and be re-interpreted in a high-level context.