This paper revisits Bloom’s 2 Sigma Problem in the context of generative AI, exploring AI agents developed through the Mollick & Mollick prompt engineering model that can serve as effective, personalized tutors in educational settings. Bloom’s original work demonstrated that students receiving individualized tutoring could outperform 2 sigma above the control students in traditional classrooms. Today, with the advent of large language models (LLMs) and generative AI (genAI), there is renewed potential to recreate Bloom’s learning experience at scale.

Departing from Mollick & Mollick's prompt engineering model, this study constructs complex, targeted prompts that transform general-purpose AI tools into domain-specific training agents. The model emphasizes critical elements such as roles, goals, constraints, pedagogical strategies, step-by-step guidance, and feedback mechanisms. By applying these principles, the resulting AI agents are designed not merely as interactive chatbots, but as specialized tutors that provide tailored feedback, scaffold knowledge acquisition, and encourage active learning.

While the integration of AI agents into educational contexts promises significant benefits, it also introduces challenges such as cognitive laziness, where students may over-rely on AI-generated responses. To address these concerns, this paper discusses strategies that incorporate metacognitive prompts and reflective activities, ensuring that students remain engaged and develop independent problem-solving skills.

Two classroom experiences are presented to illustrate the approach. The first was implemented with first-year students in a professional degree program, within the course Applied Creative Workshop for Young Teenagers. In this setting, AI agents were used to support students in designing creative communication projects for youth, offering personalized guidance and modeling iterative thinking. The second experience involved master’s students in digital communication, who used AI agents to assist in project planning, strategic analysis, and content development optimized for search engines. In both cases, the agents were created for students using the Mollick & Mollick model and tailored to the specific knowledge units of each course.

Preliminary findings from these implementations indicate that AI agents can significantly enhance learning outcomes by simulating the benefits of one-on-one tutoring. Furthermore, the results underscore the importance of prompt design in aligning AI responses with educational objectives, while also mitigating potential issues of over-dependence.

This study contributes to the growing body of research on AI in education by offering a concrete methodology for prompt engineering based on established pedagogical theory, and by providing initial empirical insights from two different educational contexts. The implications of these preliminary findings are discussed.

Keywords: GenAI, chatGPT, AI Agents, Artificial Inteligence.