A major challenge in agent-based modelling is the management of the process to generate executable simulations from the initial conceptual models. This process is complex and usually involves several roles, which may raise communication problems due to the diverse backgrounds and perspectives of participants and the use of non-explicit knowledge. This situation demands a clear separation and precise definition of the multiple aspects of the process, in order to facilitate their understanding, grasp their relationships and develop them. This paper addresses this goal with a fine-step refinement process for information based on the use of domain-specific languages. It considers analysis contexts that include a particular theoretical framework, domain, type of problem and target platform. For a given context, the process formally defines modelling languages conceptually close to the different aspects relevant to it. It also defines mappings between concepts in those languages. Researchers develop simulations by specifying models with the languages, and share and refine information by using mappings between these models. This infrastructure provides guidance throughout the process and makes the information involved explicit. A case study of continuous double auctions illustrates the approach.