Agents in many complex applications are often dynamic, which means that their abilities, services and even design objectives can change at run-time. Though a great number of agent-oriented modeling languages and methodologies have been proposed, it is still a challenge to develop complex multi-agent systems with dynamic agents. In this paper, the technology requirements to analyze, design and implement dynamic agents are discussed. A dynamic binding mechanism to develop multi-agent systems is proposed based on caste that is the basic abstraction to specify agents? behaviors and elementary modular unit to design and implement multi-agent systems. In our approach, the dynamic behaviors of agents are realized as the change of castes that agents bind to by executing "join' and "quit" operations on agent?s casteships at run-time. The mechanism also enables agents to change the status of its castes to be either active or inactive at run-time by executing operations 'activate" and "deactivate". In order to specify and analyze the adaptive agents, a graphical notation of caste transition diagram is designed and an example is illustrated. The posed problems and future research about dynamic binding mechanism are also discussed at last.