For complex systems with a large number of interacting agents, scalable models (whose computational cost does not depend on the number of particles) are attractive. Examples of such models are kinetic and fluid models, which describe the system by means of continuous quantities such as average densities. The question of rigorously linking agent-based to kinetic or fluid models is therefore crucial. In a first part of the talk, we will propose the derivation of a hydrodynamic model from an agent-based model describing consensus formation. In a second part, the validity of this model in view of the chaos assumption which states that the particles are required to be nearly independent will be discussed. A few illustrative examples in which the chaos assumption is not true will be given.