The intelligence agent architecture widely employs methods of logic of belief. The goal of the paper is to find a correct and effective inference mechanism that can substantially improve resolution based traditional methods. The semantics of the mechanism is based on Minsky's frames. Each agent is modeled by Minsky's frames with their slots representing what agent believes in. Inference process is realized by daemons filling the frames slots. The filling in this context means setting unknown slot values. The order of reasoning is established by a directed acyclic graph and driven by the topological sorting as a reasoning strategy. The inference algorithm analysis shows that the new method works in polynomial time. Therefore it is more efficient than NP, resolution based traditional methods. The correctness of object oriented implementation of the algorithm is established by considering the inference process in terms of abstract relational systems and their isomorphisms. Finally an implementation methodology of agents and their inference process in object oriented language is presented. All the considered concepts and methodology are illustrated in object oriented solution to `three wisemen problem' implemented in Smalltalk.