Twelve Angry Agents

Run your decisions through a jury of 12 AI minds before you commit.

Most people make important decisions alone or with one perspective. Twelve Angry Agents runs your topic through 12 AI personalities — a skeptic, an optimist, a devil's advocate, an ethicist, and 8 more — debating until they reach a unanimous verdict. Inspired by 12 Angry Men.

Fully local. Fully private. No API costs.

Requirements

• Python 3.11+
• Ollama installed and running

Installation

1. Install Ollama

brew install ollama      # macOS
# or download from https://ollama.com for Linux/Windows

2. Start the server and pull the model

ollama serve             # keep this running in the background

Then in a separate terminal:

ollama pull gemma4:e2b   # ~5 GB — default
ollama pull gemma4:e4b   # ~9 GB — higher quality, optional

3. Install Twelve Angry Agents in a virtual environment

python -m venv .venv
source .venv/bin/activate
pip install twelve-angry-agents

Usage

# Career decision with context
taa "I've been offered a senior engineer role at an early-stage startup — 30% pay increase, significant equity, but no runway visibility beyond 18 months. I have a mortgage and a family. Should I take it?"

# Architecture decision
taa "We're considering breaking our Django monolith into microservices. The team is 6 engineers, we deploy twice a week, and our main pain point is that a bug in the payments module blocks unrelated features from shipping."

# Pipe in a longer document
cat business_plan.txt | taa

# Use the larger model for higher quality
taa --model gemma4:e4b "We're debating whether to rewrite our mobile app in React Native or keep separate iOS and Android codebases. We have 2 mobile engineers and ship features every 2 weeks."

# Use custom agent personalities
taa --agents my_agents.yaml "Should we pivot from B2C to B2B?"

# Save the full debate transcript to a file
taa --output debate.txt "Should I accept this acquisition offer?"

How It Works

1. The Foreman gathers any missing context with 2-3 targeted questions
2. All 12 agents cast a blind vote — no peer influence
3. If split: agents debate sequentially, each reading the full argument history
4. After each round, all agents re-vote
5. Repeat until unanimous (or a hung jury after 50 rounds)
6. The Foreman delivers a prose verdict with the decisive arguments and key risks

The Jury

Customization

Edit config/agents.yaml to change any agent's personality or system prompt. The config validates that exactly 12 agents are defined.

Edit config/config.yaml to change the model, temperature, or debate limits.

License

CC BY-NC 4.0 — free for non-commercial use.