In this paper we propose an evolutionary approach to aggregate and control multiple Learning Classifier Systems (LCS) within a tree architecture. Our approach relies on two main principles. First, to base the tree control flow on a metaphor of a classifier attribute - strength, taking it as an expression of the classifier system excitement at a given time step. The tree control mechanism takes the excitement level of standard classifier systems to feed higher-level coordinator classifier systems, which will become responsible for choosing the appropriate host agent behavior. The second principle consists in relying on evolution to be the judge of the suitability of LCS aggregation. We believe that a “running time” aggregation mechanism will be useless if it is not provided a method to assess the suitability of the resulting structure. In the approach we propose, this role is played by simulated evolution of synthetic LCS based agents. The test-bed of our claims was Saavana, an Artificial Life environment modeled after a natural ecosystem where synthetic LCS based antelopes were subjected to simulated evolution. The preliminary results showed us that this approach improves the progressive adaptation of agent populations to the environment they are facing and looks promising regarding the emergence of high-level agents capable of dealing with multi-goal tasks.