A number of logical languages have been proposed to represent the dynamics of the world. Among these languages, the Situation Calculus (McCarthy and Hayes 1969) has gained great popularity. The GOLOG programming language (Levesque et al. 1997, Giacomo et al. 2000) has been proposed as a high-level agent programming language whose semantics is based on the Situation Calculus. For efficiency reasons, high-level agent programming privileges programs over plans; therefore, GOLOG programs do not consider planning. This article presents algorithms that generate conditional GOLOG programs in a Situation Calculus extended with uncertainty of the effects of actions and complete observability of the world. Planning for contingencies is accomplished through two kinds of plan refinement techniques. The refinement process successively increments the probability of achievement of candidate plans. Plans with loops are generated under certain conditions.