In this paper, we present a novel framework for asynchronous Web-based training. The proposed system has two distinguishing features. Firstly, it is based on P2P architecture for scalability and robustness. Secondly, all contents in the system are not only data but also agents so that they can mark user’s answers, can tell the correct answers, and can show some extra information without human instruction. We also present a prototype implementation of the proposed system on Maglog. Maglog is a Prolog-based framework for building mobile multi-agent systems we have developed. The user interface program of the proposed system is built on Squeak. Performance simulations demonstrate the effectiveness of the proposed system.