Digital Twins in Agriculture: From Technological Promise to Epistemological Tension in Complex Agroecosystems

Digital twin (DT) technology is attracting increasing interest as a potentially valuable tool for the future of agriculture. By offering a dynamic virtual representation of real agricultural systems, it opens up new possibilities for real-time monitoring, simulation, and decision support. In principle, such approaches could improve predictive capacity, optimize resource use, and support more responsive management strategies. However, agriculture cannot be treated as an engineered system, and this is where important challenges emerge. Agroecosystems are living, context-dependent, and inherently variable, shaped by diverse processes that remain only partly observable and often difficult to model. This makes their representation and prediction considerably more complex than in many industrial applications. In this review, we critically examine the conceptual foundations, architectural frameworks, and current applications of agricultural digital twins (ADTs), while also identifying key scientific and practical constraints that continue to limit their development. Particular attention is given to two recurring issues: the assumption that increasing data availability necessarily improves prediction, and the persistent gap between observable variables and the underlying biological and ecological processes that govern system behaviour. Drawing on conceptual figures and comparative analyses, we highlight important research gaps and argue for a shift in perspective. Rather than pursuing increasingly precise predictions, there is a need to develop digital twins that explicitly account for uncertainty and support more resilient forms of decision-making. In this context, the value of ADTs may lie less in predictive accuracy alone, and more in their ability to help decision-makers navigate complexity, variability, and change.