The architecture of an agrarian intelligent system is proposed, which forms the basis for a selflearning management decision support system. The system is designed to cover all stages of the preliminary analysis – from the agricultural problem formulation to the provision of an analytical report, forecast or recommendation. Based on the knowledge generated by the system, a person who does not even have a special education in agriculture can make an adequate managerial decision. The system consists of the following set of modules and blocks: the space of agricultural tasks, the space of data sources, data storage, journals, the space of models, the documentation space of decision support, the task (as an element of space), formalization of user data, formation of an input data array for applying the model, the model output data, indicators, models, the access to journals, data selection, the active circuit of the agrarian intelligent system, nodes of the agrarian intelligent system. In the future this system will be able to automate the mаnаgement of agricultural processes within the framework of the approach referred to as “Smart farming”. It is also proposed to use, in addition to the well-known models (imitation, optimization, and others), the concept of agent modeling, on which many modern foreign systems of predictive technologies in agriculture are based. The fl exibility of the system allows one to adapt it in order to solve the widest range of agricultural producer problems depending on the enterprise production specialization, climatic conditions of agricultural activities, the choice of cultivated crops and the level of intensifi cation of agricultural technologies. The system is built as fl exible and wide as possible in order to adapt to various requests, including those that may arise in the future, but have not yet been formulated at present.