Multiagent AI Teaching Assistant for Personalization and Adaptive Programming Education

Intelligent tutoring systems have long used learner models and instructional rules to tailor feedback. Adaptive learning platforms use quiz performance, activity history, or skill estimates to recommend the next exercise. Automated programming assessment can execute code against test cases, detect common errors, and provide rubric-based feedback. Large language models can also generate explanations and examples, but may produce inaccurate guidance or code that appears plausible without being correct. A multiagent design adds explicit coordination between specialized roles. Its value depends on whether this division improves feedback quality and adaptation enough to justify added complexity, latency, and cost. The proposed system therefore includes validation and teacher oversight as core components rather than optional add-ons.

The proposed architecture contains a learner interface, coordinator, specialized agents, learning-content repository, learner model, code execution sandbox, and instructor dashboard. The coordinator receives a learner action and current profile, routes tasks to suitable agents, checks whether the resulting recommendation satisfies pedagogical and safety constraints, and returns the response to the learner. Performance evidence is then used to update the learner model. Coordinator Agent Router Tutor Agent Concepts Code Agent Examples Assessment Quizzes Feedback Hints Profile Learner model Content Lessons Multiagent Personalization Layer

The system follows a cycle of observing, estimating, selecting, teaching, assessing, and updating. It begins by collecting relevant learning goals and recent performance. The learner model estimates which skills appear mastered and which require more practice. The assessment agent selects an activity, and the tutor or feedback agent provides support. After the learner submits an answer or code, automated checks and agent analysis generate feedback. The learner profile is updated with uncertainty retained where evidence is limited. Learner Code / quiz Profile Skill model Planner Next activity Learning Content + hints Assessment Evaluate code Performance data loops back to update the learner profile

Personalization may adjust the explanation level, example context, practice difficulty, hint depth, and time allotted. It should also offer learners choices: they may request a different explanation, skip a familiar example, or ask for a challenge. Such controls make adaptation more transparent and reduce the risk that an incorrect profile estimate traps a learner in overly simple exercises.

A teaching assistant should support learning rather than silently complete graded work. Course settings may define when full solutions are available, how hints are logged, and whether assistance must be disclosed. The system should encourage learners to explain their reasoning and modify examples themselves. Instructors should be able to configure the level of assistance for formative practice, homework, and examinations.

The system should support secure sign-in, topic selection, learner profile updates, code submission, sandboxed execution, adaptive task recommendations, staged hints, progress history, instructor review, and feedback reporting. Learners should be able to inspect and correct relevant profile information, while instructors should be able to edit learning goals and override recommendations when necessary.

A practical prototype can be developed in stages. First, create a learning interface and structured repository of topics and exercises. Second, implement deterministic code execution and test reporting in a restricted sandbox. Third, add a learner profile with transparent rules for topic recommendations. Fourth, introduce specialized agents behind a coordinator and require validation before showing feedback. Finally, add an instructor dashboard, privacy controls, audit logs, and evaluation instruments. A staged design makes it easier to identify whether each additional agent improves the learning experience.

Future development may include multilingual explanations, visual program tracing, speech-based interaction, teacher-configurable rubrics, learning analytics, and integrations with learning-management systems. Additional research could investigate agent coordination strategies, methods for detecting incorrect explanations, transparent mastery estimates, and ways to balance personalized assistance with independent practice. A classroom deployment should be iterative and guided by educators and learners. 12. Conclusion This paper proposed a Multiagent AI Teaching Assistant for Personalization and Adaptive Programming Education. The design coordinates tutor, code analysis, assessment, feedback, and learner-modelling agents through a central coordinator. A feedback loop uses learner performance to adapt examples, hints, and task difficulty, while deterministic checks and instructor oversight help manage the limitations of generated feedback. The approach may make programming support more timely and responsive, but its value must be established through careful evaluation of learning outcomes, safety, fairness, and total instructional effort. Multiagent AI should complement educators and encourage independent reasoning rather than replace teaching judgment or provide unreviewed answers.