The web size is increasing continuously. The more the Internet is growing; the more tendencies the people have to use the search engines. Moreover, since most of the commercial search engines are based on keyword indexing, there are many records in their result lists that are irrelevant to the user’s information needs. It is shown that for retrieving more relevant and precise results, the following two points should be concerned: First of all, the QUERY (either it is generated by a human or an intelligent agent) should be expressed in an accurate and exact manner. Second, we should empower search engines with the ability to capture the semantic relation between the words and the QUERY context. Hence, different search engine architectures, each of which containing QUERY REFINEMENT or semantic understanding components, have been proposed. Each architectural model has its own specific properties; but, most of them focus on only one of the two points mentioned above to improve the overall system efficiency. Moreover, in existing architectures, QUERY REFINEMENT components have direct interaction with users which may either take their time or threat their privacy while gathering basic information. In this paper, we proposed an improved architectural model for agent and ontology based search engine which uses domain ontology for semantic understanding and a QUERY REFINEMENT subsystem based on fuzzy ontology. This subsystem helps Search Agents to refine their queries, express them in a more precise way and get more relevant results. The simulation result shows that using this QUERY REFINEMENT subsystem by Search Agents can improve the system efficiency up to 5.2%.