Multi-Agent Based Simulation (MABS) is a suitable way to simulate the dynamics of natural or social systems. However, observing and analysing large-scale simulations involving numerous, complex and heterogeneous agents is still a challenge. Understanding simulation results requires the user to perform the following visual information-seeking tasks: (i) overview, (ii) zoom-and-filter, (iii) details-on-demand. This paper concentrates on the analysis of conversations among agents in MABS at various scales, from global overview to local details. Tools do exist in today’s platforms to visualise the organisational structure of a MABS (i.e., agents’ representation within their environment) and also to provide details based on local message exchanges as in the AUML formalism. However, to our knowledge, tools for viewing information exchanges among agents at various scales (namely at a macro level) are still lacking. The need to understand phenomena ranging from macro level to local interactions arise at every phase of the modelling cycle in order to answer the questions that should be asked: • Conceptual validity: Is the representation reasonable for the model’s intended use? • Computerized model verification: Does the model generate correct outputs? • Operational validation: Does the model meet the end-user needs and criteria? This paper presents both conceptual and pragmatic findings to overcome the lack of suitable representation of interactions in large-scale MABS. We introduce the concept of ‘conversation’ as a set of sub-conversations, metadata and knowledge on conversations, and optional messages identifiers. Conversation can thus be the basis for new interpretation tools more suitable for overviewing large MABS than tools based on message enumeration. Indeed, information is presented at a higher level and is more condensed. We thus propose the following tools based on conversations: (i) Conversation maps, to give an overview of all conversations that occur in a simulation; (ii) Conversation wires, to give conversation views at intermediate scales; (iii) Conversation sequence diagrams, giving the chronological details of all elementary messages of a conversation. These tools can be customized according to the user’s criteria. We illustrate the use of such concepts and tools on a real simulation experiment: the organization of organic matter transfers at a territory scale in the Reunion Island simulated with an agent-based model called Biomas. These concepts and tools for overviewing the interaction among agents are useful at every phase of the modelling cycle as they help explore simulation outputs at various scales. In research and learning activities they help identify more precisely unsatisfactory situations (e.g., failure of negotiations between agents) and improve simulation scenarios. They thus provide facilities for analysing MABS at various levels, allowing one to explain macro-level phenomena in terms of local underlying agents’ interactions. Our approach is thus a step towards answering the two issues inherent in bottom-up approaches as in MABS: identify global phenomena and local interactions accounting for them.