In this paper we present a �progressive filtering� technique aimed at improving the performances of a multiagent system devised to perform text categorization. The technique exploits the discriminant capabilities of multiple classifiers organized into a taxonomy and is aimed at coping with a problem that occurs very often in text categorization tasks, i.e. with the unbalance �for any category� between relevant and non relevant inputs. Experiments, performed on the RCV1-v2 benchmark, highlight the validity of the approach.