Multi-Agent AI System for ARC-AGI-3 Agent Preview Competition

TOMAS (Thinking, Observing, Modeling, and Acting System) mimics human cognitive processes through three specialized AI agents: perception, learning, and strategic decision-making.

🏞️ AISTHESIS - "What changed?"

• Analyzes game frames with mathematical precision
• Provides movement vectors, spatial relationships, clickable coordinates
• Always delivers precise coordinates for LOGOS decisions

🧠 SOPHIA - "What are the rules?"

• Rapidly generates and tests game mechanic hypotheses
• Rule Consolidation: "Etches successful patterns into memory" when levels complete
• 3 pathways to promote theories into confirmed rules

⚡ LOGOS - "What should I do?"

• Makes strategic decisions using human-like psychology
• 5 mental states: Exploring, Pattern-seeking, Testing, Optimization, Frustrated
• Chooses 1-5 action sequences based on confidence and emotional state

Game State → AISTHESIS → SOPHIA → LOGOS → Actions → Game State

1. 🏞️ AISTHESIS analyzes visual changes with mathematical precision
2. 🧠 SOPHIA learns patterns and consolidates successful rules
3. ⚡ LOGOS decides next actions using human-like psychology
4. 🎯 Actions execute and cycle repeats

• Rule Consolidation: Successful patterns become permanent knowledge
• Human Psychology: AI experiences frustration, curiosity, and confidence
• Mathematical Analysis: Precise movement vectors and spatial relationships
• Persistent Learning: Knowledge transfers across levels

# Run TOMAS Engine on all 3 competition games simultaneously  
uv run main.py --agent=tomasengine

Configuration: 60 moves per game (optimized for ARC-AGI-3 leaderboard performance)
Requirements: Google Gemini API, UV package manager

🧠 TOMAS Engine - Where AI meets human cognition for the ARC-AGI-3 Challenge