Increasingly complex AI systems are now often faced with the use of a combination of various theories, subsystems or architectures, whose behavior proposals have to be integrated into one behavior generation mechanism. It is the aim of this paper to present such a behavior integration mechanism. It is then evaluated on a number of small scenarios, with the idea of comparing some single "input behaviors" with the combined "output behavior" in terms of consistency, richness and credibility. The goal of this mechanism is indeed to at least preserve the quality of the inputs, and whenever possible to produce a behavior that is judged more convincing than any of its inputs taken in isolation. Applications are many, but we sketch in a concluding section the current deployment of this mechanism in a collaborative project for urban simulation and its central unifying role in its agent architecture.