Numerical methods related to reinforcement learning or bayesian inference have demonstrated impressive performances, solving some difficult tasks of decision making. Yet some ingredients must be reconsidered, as one aims at defining a cognitive agent demonstrating executive control in ecological conditions. Three fundamental dimensions are underestimated. The agent must be autonomous, cumulating its experience throughout time and with a proper motivation to act. The environment might display characteristics like volatility and stochasticity, with no preliminary segmentation of information. Its computational power complexity must be bounded in space and time. Inspired by knowledge and data from neuroscience and cognitive science, we develop a bio-inspired cognitive architecture taking a more realistic account of these dimensions. We model, in the orbital and medial regions of the frontal cortex, a series of territories representing subtle mappings of stimuli preferences and costs of actions. They are exploited for emotional and motivational control, sliding from respondent to operant representations. This analysis of internal goals and needs is used by our model of lateral prefrontal cortex to bias the default reactive behavior learned to answer external stimuli. Our model is assessed in both classical tasks and more realistic navigation in a virtual environment.