Far beyond its relevance for commercial and political marketings, opinion formation and decision making processes are central for representative democracy, government functioning, and state organization. In the present report, a stochastic agent-based model is investigated. The model assumes that bounded confidence and homophily mechanisms drive both opinion dynamics and social network evolution through either rewiring or breakage of social contacts. In addition to the classical transition from global consensus to opinion polarization, our main findings are (i) a cascade of fragmentation of the social network into echo chambers (modules) holding distinct opinions and rupture of the bridges interconnecting these modules as the tolerance for opinion differences increases. There are multiple surviving opinions associated to these modules within which consensus is formed; and (ii) the adaptive social network exhibits a hysteresis-like behavior characterized by irreversible changes in its topology as the opinion tolerance cycles from radicalization towards consensus and backward to radicalization.