Considered archaic and suffering from a "technical blockage", Roman agriculture has long been studied by historians solely through Latin agronomic treatises. This reflection on the nature and performance of this agriculture has benefited in recent decades from the contribution of numerous archaeological, bioarchaeological (archaeobotany, archaeozoology), and paleoclimatic data, which make the reflection more complex. Modelling now offers the possibility of taking into account these multidisciplinary data and testing multiple hypotheses. In this paper, we will present the design and results of a multi-agent model simulating the harvests and agricultural surpluses of villae and Gallo-Roman farms for different types of crops (wheat, legumes, vines), according to agricultural systems (techniques, tools, yields,...) and climate (temperatures, precipitation,...). The impact of the Roman Climate Optimum on agricultural yields will be measured using the Lund-Potsdam-Jena-management-Land (LPJmL) agrosystemic modelling, whose principles and operation have been integrated into the multi-agent system. We will see how the simulations make it possible to readjust the values of the different variables taken into account and to reassess the performance and productivity of agriculture in the south of France (Provence, Languedoc) between the Iron Ages and the Roman period. The results of these simulations highlight an ancient agriculture that was efficient and largely capable of generating significant surpluses in agricultural production.