AI Political Agents Project

A system for simulating political debates and negotiations between AI agents modeled after historical figures with opposing ideologies.

Research Question

Can political AI agents of historical figures who opposed each other reach a consensus in a simulated setting?

Example Scenarios

• Hitler vs Gandhi vs Jinnah: Exploring ideological conflicts and potential common ground
• US vs Japan: Could the atomic bomb have been prevented? How?
• Trump vs Mao: Trade and tariff negotiations
• Winston Churchill vs Karl Marx vs Niccolò Machiavelli: Different political ideologies

Key Features

• Historical figure personality modeling
• Multi-agent debate system
• Consensus detection and analysis
• Real-time negotiation simulation
• Optional Ollama LLM integration (local)
• Minimal FastAPI web UI

Project Structure

├── agents/           # Historical figure AI agents
├── debates/          # Debate simulation system
├── frontend/         # FastAPI + HTML minimal frontend
├── utils/            # Utilities (Ollama client)
├── consensus/        # Consensus detection algorithms
├── data/             # Historical data and context
└── examples/         # Example scenarios and outputs

Quick Start (CLI)

python3 main.py --agents hitler gandhi jinnah --topic territorial_disputes --rounds 10 --format json --summary-only

Optional: Use Ollama for LLM Responses

• Install Ollama: see https://ollama.com
• Start server (macOS): ollama serve
• Pull a model (example): ollama pull llama3.1:8b
• Configure (optional):
  • export OLLAMA_BASE_URL=http://localhost:11434
  • export OLLAMA_MODEL=llama3.1:8b
• In FastAPI UI, check "Use Ollama" and optionally set base URL/model.

Chat API (Serverless)

• Vercel route: POST /api/chat
• Required env: GEMINI_API_KEY
• Optional env: GEMINI_MODEL (default gemini-2.0-flash), GEMINI_BASE_URL
• Optional tuning:
  • CHAT_CACHE_TTL_S (default 120 seconds)
  • CHAT_CACHE_MAX_ITEMS (default 256)

Conversation Memory (Chat Continuity)

• The simulator keeps a short-term memory window (default: 8 turns) plus a rolling summary.
• It tracks salient user preferences, open loops, and consensus metrics to keep replies coherent.
• Tuning knobs (see utils/conversation_state.py):
  • memory_window (default 8)
  • compromise_threshold (default 0.7)
  • weights for consensus scoring

Run Local Frontend (FastAPI)

uvicorn frontend.server:app --reload

Then open http://127.0.0.1:8000.

Streamlit UI (Alternative)

pip3 install -r requirements.txt
streamlit run web_app.py

License

MIT