The control structure for the community of agents is not hierarchical nor is planning centralized (Georgeff, 1983). Agents do not share an overarching plan (Durfee & Lesser, 1987), rather cooperative plans emerge from local interactions (Mataric, 1992). The capacities of other agents is not treated as a given (Genesereth, Ginsberg, & Rosenschien, 1986), but this information can be implicitly acquired through the mechanisms of collective memory.
The basic cycle for the system goes as follows:
Agents have identities such that different combinations, or teams, of agents can be used for each cycle of the system. Agents can request advice from an external user during the problem-solving episode (although this is not necessary for the current set of problems).
Improvement in the performance of a community of agents occurs over several generations of problems. When solving initially, agents are satisfied with any solution. However, when preparing the solution for storage, attempts are made to improve the solution that was generated on the fly. In future episodes, fragments of the improved solution may be recalled and further improvements can be made. Thus, through the mechanisms of knowledge preparation and collective memory, the behavior of the community will evolve and improve for a certain class of problems.