We describe a hierarchical learning approach for effective coordination in repeated games based on a common-sense decomposition of the “coordination problem”. In contrast to most other research on mechanism design and game-learning, we concentrate on breaking down the top-level problem into simpler learning tasks concerned with learning (i) utility functions, (ii) best-response strategies and (iii) cooperation potentials. We also report on empirical results with the layered learning architecture LAYLA that is constructed using these sub-components in a resource-load balancing scenario. The positive results show that the approach deserves further investigation, although a number of (possibly problem-inherent) difficulties illustrate the limitations of learning approaches in real-world applications.