When we use a search engine in Internet, we often cannot retrieve a document we look for, due to the difficulty of finding adequate keywords. One of the ways to solve the problem is to search adequate keywords interacting with a user: a metasearch engine displays keywords and pages, a user evaluate them. Iterating the process, the metasearch engine narrows the keywords, getting closer to what a user looks for. For realizing it, we construct a graph called a linked document space, in which nodes and links represent the pages and the similarities, respectively. Then, the metasearch engine searches every promising page considering relations between pages, following the evaluation from a user. This paper presents a search method in the linked document space, called the socially topology agents (STA), which are inspired by human social relation and swarm intelligence. STA is applied to a multi-peak large-scale search problem, and it is shown that STA can find most of peaks quickly, as well as the aggregation and distribution of the agents in the space are controlled by a user command. They mean STA is feasible as the metasearch engine