The Silvopastoral Project is being implemented by CATIE in the Atlantic humid lowland of Costa Rica, in order to develop alternatives to the current destructive mode of livestock production. The overall approach is briefly described and two superimposed on-farm experiments are discussed in more detail. The main constraints for on-farm research within the project are analysed: the limited previous knowledge of the area and of the local species, the difficulty of achieving a common understanding about research with collaborating farmers, the need for additional staff supervision and the complexity of the statistical analyses. An attempt is made to draw lessons from the ongoing project regarding how to resolve the apparent contradiction between rigorous scientific requirements and questions of inmediate value to extension, the compromises to be achieved between different types of evaluation (statistical, risk, etc.) and the potentials and limitations of farmers' participation in research.