Marathon Agent: Autonomous Multi-Day Learning System

Marathon Agent: An autonomous multi-day learning system that plans, executes, and evaluates tasks, simulating AI reasoning to showcase full self-managing workflow.

Comment

💡 Inspiration

Education is often treated like a sprint—last-minute cramming—rather than a marathon of steady growth. I noticed that most AI tutors today are passive; they sit and wait for the student to ask questions.

I wanted to build something proactive.

My inspiration came from the idea of a "Thinking Tutor": an autonomous system that plans a curriculum, teaches actively, evaluates understanding in real-time, and slows down when a student struggles—just like a human teacher would.

🧠 What It Does

Marathon Agent is an autonomous, self-healing educational system powered by Google Gemini 3.

Goal Setting:The user provides a learning goal (e.g., "Learn Object-Oriented Programming").

Autonomous Planning:A Planner Agent generates a dynamic, multi-step syllabus.

Active Teaching:An Executor Agent teaches each lesson with explanations, analogies, and code examples.

Self-Evaluation (The Critic):A built-in Evaluator Agent grades each lesson from 0–100.

Self-Healing Curriculum: If a lesson scores below 70, the system automatically inserts a targeted "Remedial Lesson" before moving forward.

This creates a continuous feedback loop where the system adapts to the learner instead of pushing them ahead blindly.

🛠️ How I Built It

I structured the project as a Monorepo, strictly separating logic from presentation.

Backend (The Brain) Built with Node.js and the google-generative-ai SDK.

Implements three autonomous agents:Planner (curriculum), Executor (content), and Evaluator (grading).

Why Gemini 3? I chose Gemini 3 Flash Preview for its ultra-low latency. It allows agent-to-agent feedback loops to happen almost instantly, without the sluggishness of older models.

Frontend (The Face) Built with React, TypeScript, and Vite (generated via Google AI Studio).

Provides a gamified dashboard with learning streaks, daily goals, and progress tracking.

The Logic Loop I implemented a custom control loop that follows this logic:

Plaintext Next Step = Advance Curriculum (if Score >= 70) Remedial Lesson (if Score < 70)

⚠️ Challenges I Faced

The "Yes-Man" Problem: The Evaluator initially scored every lesson 100/100. I had to refine the Critic’s system instructions to enforce strict academic judgment.

Security & Git History: I accidentally committed an API key early on. This was a hard lesson in security. I learned to scrub Git history, use .gitignore effectively, and move all secrets into .env files.

Infinite Agent Loops: Autonomous agents can get stuck talking to themselves. I implemented strict state control and exit conditions to prevent this.

🏆 Accomplishments I am Proud Of

Building a fully working self-healing curriculum that detects failure and fixes itself.

Achieving near-instant agent feedback enabled by Gemini 3 Flash.

Establishing a clean, professional frontend–backend architecture.

📚 What I Learned

Specialization Wins: Multiple specialized agents outperform one large, complex prompt.

Gemini 3 Precision: The model excels at following structured instructions (like returning strict JSON for grading).

Security First: Secure key management is not an afterthought; it is essential in real-world AI development.

🚀 What’s Next

Full Integration: Connecting the backend and frontend via a REST API for real-time streaming.

Long-Term Memory: Replacing local JSON storage with a vector database (e.g., Pinecone) to remember student progress over months.

Voice Mode: Adding text-to-speech for verbal explanations.

🏁 Built for the Google AI Hackathon

Share this project:

Updates