The invention relates to architecture of a computer program in order to implement a multi-agent system. The architecture enables the agents to interact with a simulation or game world on a first plane and/or with robots in the real world. Said architecture has a second and third plane. Said second plane contains an abstract representation of the simulation world on the first plane which reduces on concepts. Said third plane implements the agents of the multi-agent system. Interfaces are only arranged between the first and second plane, and between the second and third plane, not, however, between the first and the third plane. The artificial intelligence of the agents is implemented on the second and third planes such that the simulation world of the first plane can be widened, which leads to artificial intelligence. As a result, the architecture provides a KI-middleware for, for example, computer games.