This paper describes an ongoing research project in distributed, agent-based manufacturing scheduling systems which exhibit collective system behavior. A scheduling system is described which exhibits purely low-level, reactive behavior. Two types of agents comprise the collective: part agents and machine agents. The collective is a mixed-initiative system in which agents representing parts attempt to maximize part flow to the next machine and agents representing machines attempt to maximize their utilization. Agents of the collective contain only local knowledge: a machine agent schedules operations on its machine and a part agent commits available parts to machine operations. This architecture provides an opportunity to use reinforcement learning techniques to allow an agent of the collective to improve performance by learning patterns of local interactions over time. Furthermore, both agent types of the collective exhibit anytime behavior by continuously computing the next decision until it is required to commit a decision to a fellow agent. Such agent information will be of value in further research, the goal of which is to allow the dynamic collective adapts its architecture at runtime in response to unresolvable scheduling conflicts by creating temporary, higher-level agents opportunistically to handle particular situations. The architecture of the Autonomous Manufacturing Collective (AMC) is described herein.