The traditional computational and representational theory of cognition is rife with deficiencies with respect to action. To remedy this, I separate cognition into two parts: one concerned with the online activities associated with action, and one concerned with the offline activities removed from action. I then argue that online cognition is subserved by online concepts–a novel account of concepts defined by, and ontologically equivalent to, prediction machines. Prediction machines are mental artifacts, each of which is associated with a simple, non-abstract category of perception. They are capable of picking out an extension of a class of objects, as well as generating a prediction of how that object might look from a particular angle. Each corresponding online concept can therefore pick out an extension of its corresponding object through its ability to perform classification; it enables action through its ability to generate predictions. I will argue that they and online cognition avoid the deficiencies of computational and representational theories of cognition; that they are free of the constraints of language; that they provide a basis for concept acquisition; and that they provide the beginnings of a solution to the frame problem. I supplement this with a prototype implemented in a hidden Markov model. This also helps identify the two remaining issues that need to be overcome: segmentation of the visual field, and ensuring the consistency of viewing angle between different online concepts.