Workflow systems based on mobile software agents improve overall flexibility and adaptability. The design of such systems does however, require that some attention be paid to performance tuning. This paper describes a decentralized agent control and management strategy that prevents system flooding and maintains good overall system throughput. Control of the system is divided among the following three controlling entities: the Workflow Service Broker (WSB), Agent, and Agent Pool. The WSB maintains information about what services are available on the network. The Agent maintains its itinerary, current state, and travel log; this information is used in conjunction with the information from the WSB to direct routing of the Agent. The Agent Pool maintains information about the number of agents in the system, and the overall system load. We discuss the implementation of this control strategy in a workflow application called Autopilot, which is a heterogeneous text processing workflow system where the elements are of unknown complexity and size, and where the potential processing paths through the routing domain are initially unknown. We discuss performance tuning aspects of the system and offer conclusions on such issues as agent pooling, payload simplification, object reference vs. object movement, and service co-location.