Multi-agent systems for semantic web applications require efficient implementation of agent architectures without sacrificing the flexibility and the level of abstraction that agent architectures provide. In this paper, we present an agent system, called VivAce, which is implemented in Java to achieve both high efficiency and the level of abstraction provided by the BDI agent architecture. VivAce (Vivid Agent Computing Environment) has the characteristics of a vivid agent through the BDI agent model. A vivid agent is a software-controlled system whose state comprises the mental components of knowledge, perceptions, tasks, and intentions, and whose behavior is represented by means of action and reaction rules. We first identify the requirements for multi-agent systems and then present the relevant features of VivAce and experimental results.