This research explores the abilities of vector agents (VA), as an automated processing unit, for supervised image classification methods. In comparison to the existing methods, where typically the training samples are provided by the human interpreters, VAs themselves detect, identify, extract the training samples. The advantages of the VAs, such as dynamic geometric, state and neighbourhood behaviours, allow them to extract the training samples in the same way human interpreters. In this way, they can overcome the limitations of a training process, like time consuming, difficult and uncertainty, that exist in the classical methods of this type. First, the structure of VAs is formally defined. Then, VAs are applied to extract the training areas. In the next stage, VAs use a classification algorithm such as Minimum Distance or Mahalanobis Distance to classify the satellite imagery. This method has been successfully implemented and tested on a multi-spectral satellite image to classify an IKONOS satellite image.