Agent theory · October 9, 2026
Teaching Students to Steer: Self-Regulated Learning and the Design of AI Tutors
Self-regulated learning theory offers a framework for designing AI tutors that help students plan, monitor, and evaluate their own thinking. This essay explains how educators and developers can apply these principles to build agents that foster independence rather than dependence.
Self-regulated learning is a psychological framework describing how students take control of their own cognitive processes, and it matters deeply for anyone building or deploying AI tutors in classrooms. For teachers and educational researchers, understanding this theory is essential because an AI agent that simply delivers answers can undermine the very skills students need to succeed independently.
When a student sits down with an AI tutor, the interaction often looks like a simple exchange of questions and answers. The student asks for help with a math problem, and the agent provides a step-by-step solution. But beneath that surface transaction lies a complex set of cognitive behaviors. Self-regulated learning theory, as detailed in the foundational edited volume Self-Regulated Learning: Beliefs, Techniques, and Implications for Education, frames learning not as something done to a student, but as something a student actively manages. According to this work, self-regulation involves three distinct phases: forethought, performance control, and self-reflection. In the forethought phase, a learner sets goals and plans strategies. During performance control, the learner monitors their progress and adjusts their approach. In the self-reflection phase, the learner evaluates the outcome and attributes success or failure to specific causes. If an AI agent bypasses these phases by doing the planning, monitoring, and evaluating on behalf of the student, it may improve short-term task completion while eroding long-term learning capacity.
The Architecture of Conversational Support
The challenge for educational technology is figuring out how to design an AI agent that supports these regulatory phases without taking them over. Research into conversational pedagogy offers a practical path forward. The paper Intelligent Tutoring Systems by Conversation-Based Pedagogy explores how dialogue-driven AI tutors can be structured to support student learning through conversation rather than mere information delivery. By applying cognitive and pedagogical theories directly to agent-based instruction, this research demonstrates that the way a tutor speaks—and when it chooses to remain silent—shapes how a student thinks.
A conversational AI tutor designed around self-regulated learning does not immediately answer a question. Instead, during the forethought phase, the agent might prompt the student to articulate what they already know about a topic and what strategy they intend to use. When a student encounters a difficult algebraic equation, the agent might ask, “What is your first step going to be, and why?” This forces the student to engage in goal-setting and strategic planning before touching the keyboard or pencil. The agent acts as a mirror for the student’s intentions, making implicit thoughts explicit.
During the performance control phase, the conversational model becomes even more critical. As outlined in the research on conversation-based pedagogy, intelligent tutoring systems can use dialogue to help students monitor their own comprehension. If a student makes an error, a poorly designed agent simply flags the mistake and provides the correction. An agent grounded in self-regulated learning theory, however, might ask the student to explain their reasoning aloud. By requiring the student to verbalize their thought process, the AI encourages active monitoring. The student must listen to their own logic, identify where it breaks down, and adjust their approach. The AI facilitates the regulation; the student performs it.
Evidence and the Risk of Dependency
There is substantial reason to invest in this kind of careful design. A comprehensive meta-analysis published in a top educational research journal, A Meta-analysis of the Effectiveness of Intelligent Tutoring Systems on K–12 Students’ Mathematical Learning, provides empirical evidence that intelligent tutoring systems can positively impact student achievement outcomes. When deployed effectively, these systems yield measurable gains, particularly in mathematics, where procedural steps and conceptual understanding must align. However, the effectiveness documented in this meta-analysis depends heavily on how the tutoring system interacts with the learner. An AI agent that functions as a sophisticated answer key will produce different results than one that functions as a cognitive coach.
The risk of dependency is the central tension in applying self-regulated learning theory to AI. When an agent coordinates work in a classroom or assists a teacher by managing routine inquiries, it is tempting to optimize for speed and friction reduction. Teachers are busy, and an AI that instantly resolves a student’s confusion frees up instructional time. Yet, if the resolution comes at the cost of the student’s self-reflection phase, the trade-off is ultimately harmful. In the self-reflection phase, learners must evaluate their performance and make causal attributions. Did they succeed because they used a good strategy, or because the task was easy? Did they fail because they lacked knowledge, or because they rushed? An AI tutor can guide this reflection by asking targeted questions after a task is completed, prompting the student to assess their own effort and strategy rather than just looking at a final score.
For teachers working alongside AI agents, the implications are concrete. Educators must evaluate whether the tools they adopt encourage students to plan, monitor, and reflect, or whether those tools quietly assume responsibility for those cognitive tasks. The prompts an AI uses, the pauses it builds into its responses, and the questions it asks back are all architectural choices that either scaffold self-regulation or dismantle it.
Ultimately, the goal of introducing AI into education is not to create students who are highly skilled at interacting with machines. It is to create students who are highly skilled at interacting with difficult problems. Self-regulated learning theory reminds us that the ability to steer one’s own thinking is not a natural given; it is a learned behavior. If we want AI tutors to be genuinely effective, as the evidence on intelligent tutoring systems suggests they can be, we must ensure they are built to teach students how to drive, rather than simply carrying them to the destination.