AI Arena

Multi-agent competitive LLM reasoning using Backboard API.

The Game: Grid Heist

AI Arena pits 4 different LLM agents against each other in Grid Heist, a turn-based strategy game on a 9×9 grid. Each agent must navigate the board, collect treasures, open vaults, negotiate with opponents, and outmaneuver rivals to maximize their score.

Game Mechanics

Objective: Collect the most points by the end of the match (default 15 rounds).

The Board:

• 9×9 grid with various tile types
• Fully visible - no fog of war, all players can see everything
• Deterministic - same seed produces identical layouts for replayability

Tile Types:

• Treasures (+1, +2, or +3 points) - collect by standing on them
• Keys - needed to open vaults (collected like treasures)
• Vaults (+8 points) - high-value targets requiring a key to open
• Scanners - reveal information and grant small bonuses
• Traps - can be placed by players to block opponents

Player Actions (one per round):

• MOVE - navigate the grid in 4 directions
• COLLECT - pick up treasures or keys on your current tile
• OPEN_VAULT - spend a key for 8 points (if on a vault)
• SCAN - use scanner tiles for information
• SET_TRAP - place traps on adjacent tiles
• STEAL - take keys or points from adjacent players

Strategic Elements:

• Negotiation Phase - agents can propose deals, form alliances, or make threats
• Memory - each agent remembers past negotiations, betrayals, and opponent behavior
• Multi-model routing - different LLMs bring unique strategies and personalities
• Tool usage - agents can query game state, check legal actions, and propose deals

The game rewards both strategic planning (collecting keys for vaults) and social dynamics (negotiation, betrayal, cooperation). Watch as different LLM personalities emerge through their play styles!

Quick Start

Prerequisites

• Python 3.12+
• Backboard API key

Installation

pip install -r requirements.txt

Configuration

Create a .env file in the project root with your Backboard API key and settings:

# Backboard API Configuration
# Get your API key from https://app.backboard.io
BACKBOARD_API_KEY=Enter Your API Key Here

# Backboard API settings (usually don't need to change)
BACKBOARD_BASE_URL=https://app.backboard.io/api
BACKBOARD_TIMEOUT=30

# Match Configuration
DEFAULT_MATCH_ROUNDS=15
DEFAULT_MATCH_SEED=demo_1

# UI Configuration
UI_FULLSCREEN=true
UI_DEFAULT_SPEED=1.0

# Backboard Model Routing (4 different models for demo)
P1_MODEL=gpt-4
P1_PROVIDER=openai
P2_MODEL=claude-3-5-sonnet
P2_PROVIDER=anthropic
P3_MODEL=gemini-1.5-pro
P3_PROVIDER=google
P4_MODEL=gpt-3.5-turbo
P4_PROVIDER=openai

# Web Search Configuration (set to true to enable)
ENABLE_WEB_SEARCH=false
SEARCH_BUDGET_PER_AGENT=1
SEARCH_COOLDOWN_ROUNDS=3

# Safety limits
MAX_LLM_CALLS_PER_MATCH=250

Running a Match

Run a live 15-round match with 4 AI agents:

python -m ai_arena.cli run

Run a Backboard-powered match and log to SQLite:

python -m ai_arena.cli run_backboard --seed demo_1 --rounds 10

Run with custom seed and rounds:

python -m ai_arena.cli run --seed demo_1 --rounds 20

Replaying a Match

Replay a previously recorded match:

python -m ai_arena.cli replay match_12345

Replay at faster speed:

python -m ai_arena.cli replay match_12345 --speed 2.0

Demo Checklist

See DEMO-CHECKLIST.md for a full demo flow and UI controls.

Project Structure

• src/ai_arena/engine/ - Deterministic game logic (Grid Heist)
• src/ai_arena/orchestrator/ - Multi-agent orchestration with Backboard
• src/ai_arena/rag/ - Rules and strategy retrieval
• src/ai_arena/storage/ - SQLite logging and replay
• src/ai_arena/ui/ - Pygame visualization