AI Mafia

AI Mafia is an advanced simulation of the classic social deduction game "Mafia," where AI agents play as villagers and mafia members, each with unique personalities, memory, and evolving strategies. The game is designed for research, experimentation, and entertainment, showcasing multi-agent reasoning, deception, and emergent behavior.

Orchestrator-Driven Conversation

Unlike traditional round-robin or scheduler-based agent systems, AI Mafia uses a central Orchestrator to manage the flow of conversation. The Orchestrator decides which agent should speak next based on:

• Defense Priority: If an agent is accused or mentioned, they get a chance to defend themselves immediately.
• Patience Threshold: Agents who have been quiet for too long are forced to speak, ensuring everyone participates.
• Echo Chamber Detection: If the conversation becomes repetitive, the Orchestrator selects a quiet agent to break the loop.
• No Immediate Echo: Prevents the same agent from speaking twice in a row, promoting diverse viewpoints.

Advantages of the Orchestrator approach:

• More realistic, dynamic, and unpredictable conversations
• Dominant personalities emerge naturally, as in real social deduction games
• Ensures quiet agents are not left out, but doesn't force strict turn-taking
• Adapts to the flow of accusations, alliances, and suspicion

This approach leads to emergent social dynamics and more engaging, lifelike games.

Features

• Orchestrator-Driven Conversation: Central manager decides who speaks next for natural, context-aware flow.
• AI Agents with Personalities: Each agent has a unique personality profile (aggressive, analytical, charismatic, etc.) that influences their speech and strategy.
• Persistent Memory: Agents maintain a scratchpad of past games, learning from wins and losses to adapt their tactics.
• Dynamic, Realistic Dialogue: Agents analyze conversation context and respond with evidence-based reasoning.
• Voting & Elimination: After a set number of messages, agents vote to eliminate a suspect. Voting history is tracked for pattern analysis.
• Death Wills & Mafia Editing: Eliminated villagers leave cryptic wills. Mafia can edit these to obscure clues.
• Win Condition Detection: The game automatically detects and announces when villagers or mafia win.
• Transcript Saving: Full game transcripts are saved for review and analysis.
• API-Driven Language Models: Supports OpenAI and Google Generative AI for agent dialogue generation.
• Thread-Safe Engine: Uses locks to ensure safe concurrent access to shared conversation state.

Installation

1. Clone the repository:

   git clone https://github.com/aevnum/ai-mafia.git
   cd ai-mafia

2. Install dependencies:

   Ensure you have Python 3.8+ installed. Then run:

   pip install -r requirements.txt

3. Set up API keys:

   • For OpenAI: Set the OPENAI_API_KEY environment variable.
   • For Google Generative AI: Set the GOOGLE_API_KEY environment variable.
   • You can use a .env file for convenience.

4. Run the game:

   python app.py

   (Or use Streamlit if a UI is provided: streamlit run app.py)

File Structure

• agent.py — Defines the AI agent, personality, memory, and decision logic.
• game_engine.py — Orchestrates the game, conversation, voting, and win conditions.
• api_handler.py — Handles API calls to language models.
• config.py — Game and agent configuration.
• personalities.py — Defines agent personality templates.
• scratchpads/ — Persistent memory for each agent.
• transcripts/ — Saved game transcripts.
• requirements.txt — Python dependencies.

Customization

• Add or modify personalities in personalities.py.
• Adjust game parameters in config.py.
• Plug in different language models via api_handler.py.

License

MIT License. See LICENSE for details.

Acknowledgments

• Inspired by classic Mafia/Werewolf games and AI research in multi-agent systems.
• Uses OpenAI and Google Generative AI APIs for natural language generation.