Multimodal transportation is characterized by high level of uncertainty, heterogeneity of participants (rail, maritime, road transport), and technological constraints, which make traditional centralized planning methods based on static optimization models insufficiently effective. A contradiction arises between need for global coordination of supply chains and inability to promptly adapt to dynamic changes in environment without loss of flexibility. Additional complexity comes from need to coordinate decisions among independent participants of transport process, each having own objectives and constraints. The aim of the article is to develop a self-learning multi-agent model for adaptive planning of multimodal transportation based on decentralized decision-making and parametric adaptation of agent criteria. Methodological basis combines principles of agent-based modeling, elements of decision theory, and formalization of participant behavior through universal satisfaction function reflecting multi-criteria nature of transport processes. Within study, analysis of existing approaches to multimodal transportation planning, formalization of requirements for mathematical methods, and design of agent structure representing key elements of transport system and their interaction were carried out. As a result, mathematical model is proposed in which each element of transport network is represented as autonomous agent with own criteria (cost, time, resources) and local decision-making mechanism. Interaction of agents is implemented through coordination mechanism that ensures formation of consistent transportation plan without centralized control. Key element is self-learning mechanism: system adapts priorities of agents by comparing expected and actual outcomes of operations, accumulating experience to improve accuracy of subsequent decisions and robustness to external disturbances. Novelty lies in development of formalized model of self-learning multi-agent system that enables adaptation of decision-making parameters based on retrospective analysis and allows consideration of recurring patterns. Practical significance is confirmed by possibility of using developed model in intelligent logistics platforms, as well as by reducing need for complete re-planning when deviations occur and by increasing robustness of transport processes. The results demonstrate potential of multi-agent approach for efficient management of multimodal transportation under conditions of high uncertainty.