Multi-agent-based simulations (MABS) relies on modelling the behavior of individual entities and their interactions in a virtual environment. Nowadays, MABS are used for studying various complex systems such as crowds, animal societies, ecosystems, traffic behaviors or the Market. So MABS are experimental research tools that contribute to our understanding of the mechanisms embedded in these complex systems. Still, studying some complex systems may require to consider millions of individuals. In such a case, the computing resources, which are required, represent a major obstacle for MABS end-users. In this respect, general-purpose computing on graphics processing units (GPGPU) is a relevant approach for addressing performance and scalability issues. However, GPU programming requires expert skills, which strongly limits both the accessibility and the re-usability of the frameworks developed using GPGPU. This paper presents MABS design guideline, dedicated to the GPU context, which allows the use of the GPU power without sacrificing the accessibility of MABS frameworks.