Purpose: This paper aims to sketch the emerging notion of auto-adaptive software when applied to e-learning software. Design/methodology/approach: The study and the implementation of the auto-adaptive architecture are based on the operational framework "ActiveTutor" that is used for teaching the topic of computer science programming in first-grade university courses. Findings: The paper describes an adaptive heuristic dedicated to the display of knowledge entities which is based on the dynamic balancing of four behavioral criteria for the low-level agents (cooperation, pedagogy, audacity, and sensibility). Practical implications: The heuristic may be used in all multi-agent systems where load balancing problems appear and where an auto-adaptive behavior is important. Originality/value: The paper presents an auto-adaptive algorithm.