Abstract: What is a good computational model that describes how humans both rapidly and flexibly generate their actions and decisions? Most models posit that there are two controllers that are balanced between, for example between a model-free and a model-based reinforcement-learning controller. Although such two-controller models explain the flexibility of our actions, it is not clear how the required balancing between two controllers can be computed rapidly. Here, using the active inference framework, I describe a context inference model that resolves this difficulty. The model postulates a single mechanism, where actions are computed using a likelihood and a context-specific prior over actions. It depends on the learned shape of this prior, whether a generated action appears to be habitual or goal-directed. We showed by simulations that this model replicates well-established experimental results. I will also briefly present a neuroimaging study, where we found evidence for context-specific adaptation of cognitive control. The two studies, together with recent work by other groups, point towards a fundamental computational principle based on context inference to generate rapid and flexible actions.