Land surface models (LSMs) describe how carbon and water fluxes react to environmental change. They are key component of climate models, yet they differ enormously. Many perform poorly, despite having many parameters. We outline a development strategy emphasizing robustness, reliability and realism, none of which is guaranteed by complexity alone. We propose multiple constraints, benchmarking and data assimilation, and representing unresolved processes stochastically, as tools in this endeavour.