Agricultural world models as an AI layer between predictive analytics and actionable decision-making in smart farming

Artificial intelligence is increasingly deployed in agriculture for yield forecasting, disease detection and environmental monitoring, yet most systems still cannot answer the operational question growers face: what would happen if tomorrow's management actions changed? This review defines agricultural world models as an emerging AI layer between predictive analytics and actionable decision-making, and synthesises how smart farming systems can move from passive prediction to auditable, action-conditioned simulation. It proposes a five-stage analytical framework—from data infrastructure through point prediction, process–data fusion, action-conditioned world models and closed-loop agents—to describe how agricultural AI systems acquire decision-simulation capability, using controlled-environment agriculture as the primary illustrative case. Agricultural world models differ from yield predictors, digital twins and process simulators in whether management actions enter as changeable variables and whether outputs support decision comparison. Strong predictors may suffer from a schedule memory trap, memorising management calendars rather than learning action–response dynamics. The review recommends evaluation based on state rollout, action perturbation tests, cross-context scope reporting and evidence-gated deployment from shadow advice to bounded actuation. World models should be treated as digital testbeds that connect data-driven learning with process-based agronomy under explicitly reported validity boundaries, not as replacements for field validation or mechanistic crop science.