Agent-Oriented programming (AOP) is inspired from the concepts and metaphors of multi-agent systems and borrows agent theory and technology to construct software systems. It represents a novel programming paradigm because its method, model, theory, and language are actually different from ones of existing mainstream programming technologies like OOP. As multi-agent system is considered as an effective technology to deal with the development of complex systems in open environment, AOP attracts many researchers and practitioners in the literatures of AI, software engineering and distributed computing. Significant progress has been made in the past twenty years. However, there are still great challenges to widely apply such a paradigm to support the development of complex systems in industry. In addition to using AI as basis, AOP should consider and borrow successful principles and practices of software engineering, especially existing programming paradigms, to promote its wide acceptance by software engineering practitioners. The aim of this paper is to give a systemic introduction of the research roadmap of AOP, investigate its state-of-the-art from a software engineering viewpoint by considering different programming levels of MAS and four research constituents of programming paradigms, including abstraction and model, mechanism and theory, language construct, and facility, supported platform. The survey intends to show the different research focuses and their changes in various stages. Moreover, the study identifies a number of issues and challenges in existing researches and prospect its future researches.