At aihorizon R&D, we've reached a milestone in developing OverMind, an experimental AI tutoring system designed to overcome the limitations of traditional prompt-response interfaces. Rather than simply reacting to isolated queries, OverMind introduces coordinated AI agents capable of planning, guiding, and adapting a complete learning experience.

By integrating agent-based workflows, session memory, and real-time learner modeling, OverMind offers a more structured and interactive approach to AI-assisted education.

 

Project Overview

OverMind's core innovation lies in shifting from reactive answers to proactive learning guidance. The system engages users in dynamic tutoring sessions that adjust to their background knowledge, learning goals, and ongoing performance.

​

Agent Roles

• Meta Agent: Designs the session strategy, monitors learner progress, and updates the plan as needed.

• Tutor Agent: Delivers instructional material, asks questions, and interacts with learners.

Together, these agents create a feedback-driven learning path that evolves continuously throughout each session.

 

Technical Architecture

Built on Azure’s AI ecosystem, OverMind is designed for modularity, extensibility, and real-time responsiveness:

• User Interface
  Captures initial learner context: goals, preferences, and prior knowledge.

• Meta Agent (LLM1)

  • Analyzes user input and session history

  • Generates and updates learning plans

  • Coordinates overall flow of instruction

• Tutor Agent (LLM2)

  • Presents content and questions

  • Engages in dialogue with the learner

  • Provides targeted, in-the-moment feedback

• Azure AI Orchestration

  • AI Agent Service: Manages and scales agent workloads

  • Semantic Kernel: Controls task flow and memory handling

  • Vector Search (optional): Supports retrieval-augmented content

• Data Infrastructure

  • Cosmos DB: Stores session state, profiles, and logs

  • Azure Monitor: Enables performance monitoring and diagnostics

 

Learning Workflow

At the heart of OverMind is a responsive agent loop that adjusts the session based on each learner’s interactions:

1. Initial Input
   The learner defines objectives and relevant background.

2. Session Planning
   The Meta Agent selects a learning path (e.g., diagnostics, concept review, exercises).

3. Instruction Delivery
   The Tutor Agent carries out the plan through content delivery and question-based engagement.

4. Response Logging
   Learner inputs are recorded and session data is updated in real time.

5. Reflection and Adaptation
   The Meta Agent reviews performance and refines the plan as needed.

6. Session Progression
   The loop continues until goals are met or the session concludes.

This structure enables OverMind to deliver instruction that feels coherent and personalized—not just responsive, but strategic.

 

Conclusion

OverMind v1 marks the creation of a functional prototype for a new class of AI tutoring systems—those that guide rather than just respond. By coordinating language model agents through orchestrated workflows and adaptive memory, OverMind lays the foundation for future AI learning platforms that are context-aware, strategy-driven, and learner-centered.

 

Contribution: Moritz Goeke (Technical Lead), Moritz Weckbecker (Conceptualization), Michael Jülich (Conceptional Lead)