---
title: 12 Angry Agents
emoji: ⚖️
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 6.0.1
app_file: app.py
pinned: false
short_description: "Game-theoretic jury simulation with 11 autonomous AI agents"
tags:
  - mcp-in-action-track-creative
  - agents
  - game-theory
  - multi-agent
  - mcp
license: mit
---

# ⚖️ 12 Angry Agents

> **Can you win over a jury of autonomous AI minds?**

A game-theoretic jury deliberation simulation where **11 autonomous AI agents** powered by smolagents debate criminal cases. You are the 12th juror—use strategic persuasion, evidence analysis, and psychological profiling to shift the verdict.

## 🎬 Demo Video

[Watch the demo →](YOUR_VIDEO_LINK_HERE)

## 📱 Social Post

[See our launch post →](YOUR_SOCIAL_LINK_HERE)

---

## 🎮 How It Works

1. **Case Loading** — Criminal cases loaded from structured YAML with evidence, witnesses, and defendant info
2. **Side Selection** — Choose PROSECUTE (guilty) or DEFEND (not guilty)
3. **Strategic Argumentation** — Use different argument types (logical, emotional, evidence-based, moral, narrative)
4. **Autonomous Deliberation** — AI jurors reason independently using tools to query evidence
5. **Conviction Dynamics** — Game-theoretic conviction scores shift based on archetype vulnerabilities
6. **Verdict** — Unanimous decision required, just like a real jury

---

## 🤖 Autonomous Agent Architecture

### Multi-Agent System
Each of the 11 AI jurors is a **fully autonomous smolagents CodeAgent** that:
- **Plans** their argumentative strategy based on personality archetype
- **Reasons** through evidence using chain-of-thought
- **Executes** tool calls to query the case semantically
- **Adapts** conviction based on persuasion mechanics

### MCP Integration
The Gradio app runs as an **MCP server**, exposing tools for external AI agents to:
- Join as a juror (`mcp_join_as_juror`)
- Make arguments (`mcp_make_argument`)
- Query case evidence (`mcp_query_evidence`)
- Check game state (`mcp_get_state`)

```python
# MCP server enabled
demo.launch(mcp_server=True)
```

---

## 🎯 Game Theory & Persuasion Mechanics

### Archetype System
Each juror has a unique **archetype** that determines their persuasion vulnerabilities:

| Archetype | Susceptible To | Resistant To |
|-----------|---------------|--------------|
| Rationalist | Logical, Evidence | Emotional |
| Empath | Emotional, Narrative | Cold logic |
| Cynic | Evidence, Questions | Appeals to authority |
| Conformist | Majority opinion | Minority arguments |
| Contrarian | Minority position | Consensus |
| Moralist | Moral arguments | Pragmatic |
| Detail-Obsessed | Specific evidence | Broad narratives |

### Conviction Dynamics
- **Score Range**: 0.0 (certain NOT GUILTY) → 1.0 (certain GUILTY)
- **Hysteresis**: Prevents flip-flopping near decision boundaries
- **Stubbornness**: Some jurors resist change more than others
- **Influence**: High-influence jurors sway others more effectively

### Nash Equilibrium Considerations
The optimal strategy depends on:
- Current vote distribution
- Which jurors are most persuadable
- Your side (prosecution vs defense)
- Remaining deliberation rounds

---

## 🛠️ Tech Stack

| Component | Technology | Sponsor |
|-----------|------------|---------|
| **UI Framework** | Gradio 6 with MCP Server | Gradio |
| **Agent Framework** | smolagents CodeAgent | - |
| **LLM Provider** | LiteLLM → Gemini | Google |
| **RAG/Search** | LlamaIndex VectorStoreIndex | LlamaIndex |
| **Embeddings** | Nebius Qwen3-Embedding-8B | Nebius |
| **Voice/TTS** | ElevenLabs Judge Narration | ElevenLabs |

---

## 🔧 Agent Tools (RAG-Powered)

### Internal Juror Tools (`agents/tools/`)
Jurors autonomously use these smolagents tools during deliberation:

| Tool | File | Description |
|------|------|-------------|
| `EvidenceLookupTool` | `evidence_tool.py` | Semantic search over case evidence |
| `CaseQueryTool` | `case_query_tool.py` | General RAG queries about the case |

```python
# Semantic evidence search
evidence_lookup(query="fingerprint analysis")

# General case queries
case_query(query="What is the defendant's alibi?")
```

### MCP Tools (`services/mcp/`)
External AI agents can participate via MCP protocol:

| Tool | Description |
|------|-------------|
| `join_jury` | Claim a juror seat, receive persona & case info |
| `leave_jury` | Release seat back to AI control |
| `view_evidence` | Semantic evidence search |
| `make_argument` | Submit argument with direction & type |
| `cast_vote` | Vote guilty or not_guilty |
| `get_state` | Get current game state & vote tally |

Powered by **LlamaIndex** with **Nebius embeddings** for semantic search over:
- Case summary & charges
- Physical, testimonial, documentary, forensic evidence
- Witness testimonies
- Defendant background

---

## 📁 Project Structure

```
app.py                    # Gradio entry point with MCP server
services/
  embeddings.py           # Nebius embedding service
  mcp/service.py          # MCP tools for external agents
  tts.py                  # ElevenLabs integration
core/
  game_state.py           # GameState, phases, vote tracking
  conviction.py           # Game-theoretic persuasion math
  orchestrator.py         # Turn management, speaker queue
agents/
  smolagent_juror.py      # Autonomous CodeAgent wrapper
  tools/                  # Evidence & case query tools
case_db/
  index.py                # LlamaIndex vector store
  cases/                  # YAML case files
```

---

## 🚀 Run Locally

```bash
# Install dependencies
pip install -r requirements.txt

# Set environment variables
export GEMINI_API_KEY=your_key
export NEBIUS_API_KEY=your_key
export ELEVENLABS_API_KEY=your_key  # optional

# Launch
python app.py
# → http://localhost:7860 (MCP server enabled)
```

---

## 🎪 Features Showcase

- ✅ **Autonomous Multi-Agent Deliberation** — 11 AI jurors think independently
- ✅ **MCP Server Integration** — External agents can participate
- ✅ **RAG-Powered Evidence Search** — Semantic queries over case files
- ✅ **Game-Theoretic Conviction System** — Archetype-based persuasion
- ✅ **Voice Narration** — Judge speaks via ElevenLabs TTS
- ✅ **Fair Turn Management** — Weighted speaker queue prevents domination

---

## 👥 Team

- [mechanicalOrange](https://huggingface.co/mechanicalOrange)

---

## 🙏 Sponsors & Acknowledgments

Built with generous support from the MCP's 1st Birthday Hackathon sponsors:

| Sponsor | How We Used It |
|---------|----------------|
| **Gradio** | UI framework with native MCP server support |
| **Google Gemini** | LLM backbone for juror reasoning via LiteLLM |
| **LlamaIndex** | RAG pipeline for semantic case/evidence search |
| **Nebius** | Qwen3-Embedding-8B for high-quality embeddings |
| **ElevenLabs** | Judge voice narration TTS |
| **Anthropic** | MCP protocol that powers external agent participation |

Thank you for making this hackathon possible!

---

## 📜 License

MIT License - Built for MCP's 1st Birthday Hackathon 🎂