Agent0 Interactive Analogy

This project provides an interactive visualization of the Agent0 framework described in the paper "Agent0: Unleashing Self-Evolving Agents from Zero Data".

• Paper: arXiv:2511.16043
• Code: GitHub Repository
• GitHub Page: GitHub Example

How to Run

Simply open the index.html file in any modern web browser.

The Analogy

The visualization uses a "Riddle Master vs. Detective" analogy to explain the complex reinforcement learning concepts:

• The Riddle Master (Curriculum Agent): Tries to create tasks that are challenging but solvable. It gets "points" (rewards) when the Detective is confused or has to use tools.
• The Detective (Executor Agent): Tries to solve the tasks. It gets "points" for correct answers. It has a "Gadget Belt" (Tools) to help with hard problems.

Key Concepts Visualized

• Co-evolution: Both agents level up together until they reach "Convergence" (Max Level).
• Tool Integration: The Detective uses tools, which forces the Master to create harder tasks.
• Uncertainty Reward: The Master benefits when the Detective is unsure.
• Failure Modes: The simulation shows what happens when tasks are "Too Easy" (No reward) or "Confusing" (Uncertainty reward only).

Simulation Controls

• Run Cycle: Manually triggers one iteration of the training loop.
• Auto Play: Runs the simulation continuously until the agents converge.
• Reset: Resets the agents to level 1.

Generated Page