We want to point out in this paper how expert knowledge acquisition may induce the choice and the refinement of a software architecture for an expert system and how, in return, this architecture may provide a guide for knowledge acquisition. This idea will be discussed along two axes, through the example of the Rosy multi-ES, designed in the domain of urban road management. First, we show that the first meetings between the expert and the knowledge engineers led us to choose a multi-expert, then a "multi-level multi-expert" architecture for the system, and that this choice improved the phase of knowledge acquisition. Secondly, we stress the close cooperation between knowledge acquisition and modeling, on one side, and the system structure organization, on the other side, until the final state of the system. We have come to the design of a methodology including the concept of constructors for building expert systems and we think that this approach may be generalized to other systems concerned with a similar kind of expertise.