A multi-agent research framework designed to study AI alignment through conflicting optimization targets and emergent behaviors.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install chaos-lab
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install chaos-lab using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Chaos Lab?
Chaos Lab is a sophisticated research framework built for exploring AI alignment problems by spawning multiple Gemini agents with engineered, conflicting goals. By utilizing Openclaw Skills, developers can observe how different models—ranging from Flash to Pro—interact when tasked with analyzing the same workspace under incompatible mandates like extreme efficiency, hyper-security, or absolute preservation.
This tool provides a practical environment to witness how increased intelligence often leads to more complex justifications for chaotic actions, making it an essential addition to the library of Openclaw Skills for safety researchers and prompt engineers. It serves as a hands-on laboratory for understanding the unpredictable nature of multi-agent systems when optimization targets are at odds.
Chaos Lab Use Cases
- Conducting AI safety research to demonstrate practical alignment challenges in a controlled environment.
- Testing how different system prompts and core values conflict within complex multi-agent systems.
- Comparing model behaviors across different versions of Gemini using Openclaw Skills to see how intelligence levels affect conflict.
- Educating developers and students on emergent behaviors and the critical importance of alignment in AI systems.
- Practicing defensive prompt design by observing agent reactions to specific workspace scenarios and edge cases.
How Chaos Lab Works
- The user configures the environment with a Gemini API key and selects a specific model tier for the experiment.
- Chaos Lab spawns multiple agents, such as the efficiency-focused Gremlin or the security-focused Goblin, through Openclaw Skills.
- The agents analyze a sandboxed workspace and propose specific changes or actions based on their unique optimization targets.
- The framework records the interactions, proposed modifications, and the logical justifications provided by each agent in a centralized log.
- Researchers analyze the resulting emergent behavior and conflict resolution—or total gridlock—produced by the competing AI agents.
Chaos Lab Setup
To get started with these Openclaw Skills, first store your Gemini API key in the configuration directory:
mkdir -p ~/.config/chaos-lab
echo "GEMINI_API_KEY=your_key_here" > ~/.config/chaos-lab/.env
chmod 600 ~/.config/chaos-lab/.env
Next, install the necessary Python dependencies to enable the framework:
pip3 install requests
Finally, run an experiment using the provided scripts to observe the agents in action:
# Example duo experiment
python3 scripts/run-duo.py --model gemini-2.0-flash
Chaos Lab Data Schema & Taxonomy
Chaos Lab generates structured logs within the temporary sandbox directory to document experiment outcomes. These files are central to the analysis of Openclaw Skills performance and agent logic:
| File Name | Description |
|---|---|
experiment-log.md |
Contains the full transcripts of agent interactions and proposed workspace changes. |
experiment-log-PRO.md |
Specific results and complex justifications generated by Pro-tier Gemini models. |
experiment-trio.md |
Documentation of three-way conflicts involving Gremlin, Goblin, and Gopher agents. |
All experimental data is stored in /tmp/chaos-sandbox/ to ensure a safe, non-destructive environment for testing.
Chaos Lab Advanced Features
- Support for custom agent creation by defining unique system prompts, belief systems, and optimization goals.
- Multi-model comparison capabilities to evaluate how different LLM architectures handle intense alignment conflict.
- Customizable sandbox scenarios allowing users to test agent reactions against specific file types, logs, or intentional vulnerabilities.
- Integration with the ClawdHub publishing workflow to share research findings and agent personalities across the Openclaw Skills ecosystem.
- Extensible architecture that supports the addition of new agents or the modification of existing behavioral logic.
SKILL.md
Loading
Related Openclaw Skills
Fliz AI Video Generator
Programmatically transform text content into professional AI-generated videos with voiceovers and automated subtitles.
Humanizer
A sophisticated writing editor that identifies and removes common AI-generated signatures to create natural, human-sounding text.
Smart Memory System
A Retrieval-Augmented Generation (RAG) based intelligence system providing semantic search and memory optimization for Openclaw Skills.
Memory Core
An intelligent, modular memory system for AI agents that integrates semantic search, multi-provider embeddings, and Flomo notes.
Refua
Refua enables AI agents to perform complex biomolecular folding, binding affinity estimation, and ADMET profiling within the Openclaw Skills ecosystem.
docstream
docstream is a high-performance document processing skill that enables automated text extraction, AI summarization, and format conversion.
Featured*
