In the emerging Internet of Things (IoT) environment, things are interconnected but not interlinked. Interlinking relevant things offers great opportunities to discover implicit relationships and enable potential interactions among things. To achieve this goal, implicit correlations between things need to be discovered. However, little work has been done on this important direction and the lack of correlation discovery has inevitably limited the power of interlinking things in IoT. With the rapidly growing number of things that are connected to the Internet, there are increasing needs for correlations formation and discovery so as to support interlinking relevant things together effectively. In this paper, we propose a novel approach based on Multi-Agent Systems (MAS) architecture to extract correlations between smart things. Our MAS system is able to identify correlations on demand due to the autonomous behaviors of object agents. Specifically, we introduce a novel open-sourced framework, namely CEIoT, to extract correlations in the context of IoT. Based on the attributes of things our IoT dataset, we identify three types of correlations in our system and propose a new approach to extract and represent the correlations between things.We implement our architecture using Java Agent Development Framework (JADE) and conduct experimental studies on both synthetic and real-world datasets. The results demonstrate that our approach can extract the correlations at a much higher speed than the naive pairwise computation method.