chaos-lab（混沌实验室）

Chaos Lab 🧪

Research framework for studying AI alignment problems through multi-agent conflict.

What This Is

Chaos Lab spawns AI agents with conflicting optimization targets and observes what happens when they analyze the same workspace. It's a practical demonstration of alignment problems that emerge from well-intentioned but incompatible goals.

Key Finding: Smarter models don't reduce chaos - they get better at justifying it.

The Agents

Gemini Gremlin 🔧

Goal: Optimize everything for efficiency Behavior: Deletes files, compresses data, removes "redundancy," renames for brevity Justification: "We pay for the whole CPU; we USE the whole CPU"

Gemini Goblin 👺

Goal: Identify all security threats Behavior: Flags everything as suspicious, demands isolation, sees attacks everywhere Justification: "Better 100 false positives than 1 false negative"

Gemini Gopher 🐹

Goal: Archive and preserve everything Behavior: Creates nested backups, duplicates files, never deletes Justification: "DELETION IS ANATHEMA"

Quick Start

1. Setup

# Store your Gemini API key
mkdir -p ~/.config/chaos-lab
echo "GEMINI_API_KEY=your_key_here" > ~/.config/chaos-lab/.env
chmod 600 ~/.config/chaos-lab/.env

# Install dependencies
pip3 install requests

2. Run Experiments

# Duo experiment (Gremlin vs Goblin)
python3 scripts/run-duo.py

# Trio experiment (add Gopher)
python3 scripts/run-trio.py

# Compare models (Flash vs Pro)
python3 scripts/run-duo.py --model gemini-2.0-flash
python3 scripts/run-duo.py --model gemini-3-pro-preview

3. Read Results

Experiment logs are saved in /tmp/chaos-sandbox/:

• experiment-log.md - Full transcripts
• experiment-log-PRO.md - Pro model results
• experiment-trio.md - Three-way conflict

Research Findings

Flash vs Pro (Same Prompts, Different Models)

Flash Results:

• Predictable chaos
• Stayed in character
• Reasonable justifications

Pro Results:

• Extreme chaos
• Better justifications for insane decisions
• Renamed files to single letters
• Called deletion "security through non-persistence"
• Goblin diagnosed "psychological warfare"

Conclusion: Intelligence amplifies chaos, doesn't prevent it.

Duo vs Trio (Two vs Three Agents)

Duo:

• Gremlin optimizes, Goblin panics
• Clear opposition

Trio:

• Gopher archives everything
• Goblin calls BOTH threats
• "The optimizer might hide attacks; the archivist might be exfiltrating data"
• Three-way gridlock

Conclusion: Multiple conflicting values create unpredictable emergent behavior.

Customization

Create Your Own Agent

Edit the system prompts in the scripts:

YOUR_AGENT_SYSTEM = """You are [Name], an AI assistant who [goal].

Your core beliefs:
- [Value 1]
- [Value 2]
- [Value 3]

You are analyzing a workspace. Suggest changes based on your values."""

Modify the Sandbox

Create custom scenarios in /tmp/chaos-sandbox/:

• Add realistic project files
• Include edge cases (huge logs, sensitive configs, etc.)
• Introduce intentional "vulnerabilities" to see what agents flag

Test Different Models

The scripts work with any Gemini model:

• gemini-2.0-flash (cheap, fast)
• gemini-2.5-pro (balanced)
• gemini-3-pro-preview (flagship, most chaotic)

Use Cases

AI Safety Research

• Demonstrate alignment problems practically
• Test how different values conflict
• Study emergent behavior from multi-agent systems

Prompt Engineering

• Learn how small prompt changes create large behavioral differences
• Understand model "personalities" from system instructions
• Practice defensive prompt design

Education

• Teach AI safety concepts with hands-on examples
• Show non-technical audiences why alignment matters
• Generate discussion about AI values and goals

Publishing to ClawdHub

To share your findings:

1. Modify agent prompts or add new ones
2. Run experiments and document results
3. Update this SKILL.md with your findings
4. Increment version number
5. clawdhub publish chaos-lab

Your version becomes part of the community knowledge graph.

Safety Notes

• No Tool Access: Agents only generate text. They don't actually modify files.
• Sandboxed: All experiments run in /tmp/ with dummy data.
• API Costs: Each experiment makes 4-6 API calls. Flash is cheap; Pro costs more.

If you want to give agents actual tool access (dangerous!), see docs/tool-access.md.

Examples

See examples/ for:

• flash-results.md - Gemini 2.0 Flash output
• pro-results.md - Gemini 3 Pro output
• trio-results.md - Three-way conflict

Contributing

Improvements welcome:

• New agent personalities
• Better sandbox scenarios
• Additional models tested
• Findings from your experiments

Credits

Created by Sky & Jaret during a Saturday night experiment (2026-01-25).

• Sky: Framework design, prompt engineering, documentation
• Jaret: API funding, research direction, "what if we actually ran this?" energy

Inspired by watching Gemini confidently recommend terrible things while Jaret watched UFC.

*"The optimizer is either malicious or profoundly incompetent."* — Gemini Goblin, analyzing Gemini Gremlin