WaterTheft develops a multi-level modeling framework to forecast adaptation surprises emerging from nonlinear behavior in human–water systems. We combine microeconomic mathematical programming models of irrigators, integrated with behavioral economics, to represent behavior under bounded rationality and identify nonlinearities – such as water theft. These individual-level models are embedded in spatial agent-based models and coupled with regional-to-global computable general equilibrium models. The integrated human system is then endogenized into spatially distributed hydrological and hydrogeological models. Ensemble experiments across models, parameters, and scenarios are used to identify uncertainty-driven nonlinearities and surprises.