GENERATIVE/DETERMINISTIC SOFTWARE ENGINEERING

PEDAGOGY WITH LOCAL MULTI-AGENT SYSTEMS MANAGING COGNITIVE/COMPUTATIONAL LOADS

Abstract

The integration of intelligent systems poses significant challenges to software development education. Current policies often overlook the complexity of emerging technologies, such as memory-efficient model fine-tuning and the orchestration of multiple local intelligent models, as well as the significant constraints imposed by limited computational resources. This study proposes a pedagogical model centered on the critical interdependence between students’ mental effort and machine processing demands. Empirical findings indicate that optimizing processing demand (e.g., through aggressive quantization) may increase mental effort, while the complexity of systems involving multiple intelligent models amplifies both cognitive and computational loads. The model seeks a pedagogical “sweet spot,” where moderate processing demand (e.g., more stable quantization) minimizes mental effort, adapting to the learner’s level of experience. It aims to develop metacognitive skills for managing these trade-offs in complex tasks, particularly under hardware constraints. Teaching strategies include progressive sequencing, comparative activities, and asynchronous execution for tasks with high processing demands, with results analyzed through reflective practices. Support tools, such as the CCLOLMAS platform (available at https://github.com/cclolmas/app), make these dynamics visible. The objective is to reshape curricula, aligning educational practices with industry demands and enabling students to effectively manage these essential trade-offs in professional software engineering practice.