Reactive multi-agent systems present global behaviours uneasily linked to their local dynamics. When it comes to controlling such a system, usual analytical tools are difficult to use so specific techniques have to be engineered. We propose an experimental dynamical approach to enhance the control of the global behaviour of a reactive multi-agent system. We use reinforcement learning tools to link global information of the system to control actions. We propose to use the behaviour of the system as this global information. The behaviour of the whole system is controlled thanks to actions at different levels instead of building the behaviours of the agents, so that the complexity of the approach does not directly depend on the number of agents. The controllability is evaluated in terms of rate of convergence towards a target behaviour. We compare the results obtained on a toy example with the usual approach of parameter setting.