Adaptive Multi-Agent Architecture to Track Students’ Self-Regulated Learning Babak Khosravifar 1 , Roger Azevedo 1 , Reza Feyzi-Behnagh 1 , Michelle Taub 1 , Gautam Biswas 2 , and John S. Kinnebrew 2 1 McGill University, The Laboratory for the Study of Metacognition and Advanced Learning Technologies, Montreal, Canada 2 Vanderbilt University, The Teachable Agents Group at Vanderbilt University, Nashville, USA (babak.khosravifar,roger.azevedo)@mcgill.ca, (reza.feyzibehnagh,michelle.taub)@mail.mcgill.ca, (gautam.biswas,john.s.kinnebrew)@vanderbilt.edu Abstract. Intelligent Tutoring Systems (ITS) can be designed to im- prove learning and performance through Pedagogical Agents (PAs) that are designed to foster self-regulated learning through interactions and ex- change of information with human learners. PAs are intelligent and follow rational behaviors, but to adaptively track students’ progress, they need to be systematically and specifically designed. However, in order to fol- low a common goal, different self-regulatory systems have been designed that use PAs, but fail to provide an adaptive multi-agent architecture which provides such feature that agents adaptively track students’ scaf- folding. In this paper, we introduce a multi-agent framework designed for an agent-based ITS. We also define the agent architecture, multi-agent framework and communication mechanism. Keyword. Pedagogical Agents, Self-Regulated Learning, Multi-Agent Systems, Agent Communication Mechanism. 1 Introduction Increasing adaptivity is being devoted to frameworks involving intelligent com- ponents that receive (or search for) data and dynamically update their internal engine to efficiently acquire and integrate information. This adaptivity is becom- ing a crucial feature in ITSs that provide scaffolding for students to effectively self-regulate their learning. There are various ITSs [1–4, 6], which are used to con- duct educational research. But in this paper, we only concentrate on agent-based ITSs [1, 3, 4, 6] where PAs continuously interact with students and objectively provide guidance to facilitate the process of learning and use of effective SRL processes. We concentrate on this category of ITSs because agents are intelligent components that could be equipped with adaptive applications and dynamically track student behaviour, based on the scaffolding they are receiving. Current ITSs are not entirely adaptive to students’ knowledge acquisition during learning in real-time. This may be because in most agent-based ITSs [1, 49