Figure 1.

The Markov blanket forms a statistical separation between an agent and its environment. Arrows indicate the “generative” direction. Conditioning on the Markov blanket renders information about the internal states of the agent independent from information about hidden state of the environment. Blanket states for the agent are the variables for controls u_t and observations x_t. The solid diamond node in the Markov blanket links posterior beliefs over the agent's control variables to environmental actions, and the solid square node in the Markov blanket passes environmental outcomes to agent observations.