Higher education students have abundant access to content, but lack individual mediation within a study plan. Conversational systems supported by Large Language Models (LLMs) have expanded personalized education at scale, but text generation does not ensure explicit pedagogical goals, progress monitoring, or the fulfillment of long-term plans. In tutoring systems, this limitation produces a reactive behavior, in which the system responds to the student without guiding them toward their learning objectives. This dissertation proposes and evaluates, following the Design Science Research approach, a multi-agent system for managing educational dialogues guided by study plans. Specialized agents supported by LLMs formulate questions calibrated according to the levels of Bloom’s Taxonomy, evaluate student responses, and conduct explanations, doubts, and session closure, while a deterministic navigation component decides which knowledge component to address and when to advance, reinforce, or revisit. Navigation operates over a prerequisite graph, adapts a computerized adaptive testing algorithm, and tracks progress through evidence propagated through the graph. With this division, the model’s variability remains confined to interpretative and generative functions, and progression becomes deterministic and auditable. The implemented prototype was verified through automated tests, behavioral invariants, and simulations. In a user study with 37 students on Python programming fundamentals, 73% perceived a meaningful sequence of topics, 70% considered the return to focus adequate, and 76% perceived progression toward an objective. The results provide evidence of feasibility for the approach in the investigated context.