Does extensive practice reduce or eliminate central interference in dual-task processing? We explored the reorganization of task architecture with practice by combining interference analysis (delays in dual-task experiment) and random-walk models of decision making (measuring the decision and non-decision contributions to RT). The main delay observed in the Psychologically Refractory Period at short stimulus onset asynchronies (SOA) values was largely unaffected by training. However, the range of SOAs over which this interference regime held diminished with learning. This was consistent with an overall shift observed in single-task performance from a highly variable decision time to a reliable (non-decision time) contribution to response time. Executive components involved in coordinating dual-task performance decreased (and became more stable) after extensive practice. The results suggest that extensive practice reduces the duration of central decision stages, but that the qualitative property of central seriality remains a structural invariant. (Contains 5 figures, 2 supplementary figures, 1 supplementary data and 4 tables.)