In this paper, a dynamic cooperation model for different agents' message exchanging is proposed. It is thought that the curve of surrounding words' relevant influence approximatively fits normal distribution with a mean of 0 and a standard deviation of σ. Following those mechanisms, an agent-based selective information retrieval on the base of topical crawling, high-performance indexing and efficient search algorithm is implemented. The agents' cooperation model expresses the procedure of agent organization and communication. A topical crawling based on anchor word concept analysis can guarantee a higher performance. The topical crawler makes it easy to find relevant pages as many as possible and irrelevant pages as few as possible.