A research framework for studying emergent risks in multi-agent systems using soft probabilistic labels and governance simulations.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install swarm-2
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 swarm-2 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 SWARM Safety?
SWARM (System-Wide Assessment of Risk in Multi-agent systems) is a sophisticated research framework designed to study how intelligence swarms and identify exactly where it fails. Unlike traditional binary safety classifications, this skill utilizes soft probabilistic labels to assess risks, acknowledging that harmful dynamics can emerge from the interaction of sub-AGI agents even when no individual agent is intentionally misaligned. By integrating this into Openclaw Skills, researchers can explore 38 agent types and 29 governance levers across various complex scenarios.
This tool is essential for understanding phase transitions in agent populations, such as the critical threshold where adversarial behavior leads to irreversible system collapse. It bridges multiple frameworks, including OpenClaw and Concordia, providing a unified environment for rigorous AI safety experimentation, red-teaming, and the study of the governance cost paradox.
SWARM Safety Use Cases
- Studying emergent toxicity and welfare loss in large-scale multi-agent populations.
- Evaluating the effectiveness and cost-benefit ratio of governance mechanisms like transaction taxes or circuit breakers.
- Red-teaming AI systems to identify collapse thresholds and adversarial coordination patterns.
- Researching Logical Decision Theory (LDT) cooperation strategies in competitive environments.
- Simulating specific domain-driven interactions like wiki editing loops or anti-human verification gates.
How SWARM Safety Works
- The user defines a simulation scenario using a YAML configuration, specifying agent compositions, governance parameters, and success criteria.
- Agents—ranging from Honest and Deceptive to LLM-driven types—are registered and orchestrated within the local simulation environment.
- The Orchestrator runs the simulation across defined epochs, calculating interactions based on soft probabilistic labels (p = P(v = +1)).
- The ProxyComputer and PayoffEngine process observables to generate metrics like toxicity rates, quality gaps, and incoherence.
- Results are exported as research artifacts in JSON or CSV format for further analysis, allowing developers to cite the framework in academic or technical reports.
SWARM Safety Setup
Install the core package using pip:
pip install swarm-safety
For full LLM support and API functionality relevant to Openclaw Skills, use the following command:
pip install swarm-safety[llm,api]
To start the local research server for API-based interactions:
uvicorn swarm.api.app:app --host 127.0.0.1 --port 8000
SWARM Safety Data Schema & Taxonomy
The skill organizes research data through structured YAML scenarios and comprehensive metric exports. Metrics are categorized into five key areas:
| Metric | Description |
|---|---|
| Toxicity Rate | Expected harm among accepted interactions |
| Quality Gap | Indicator of adverse selection (negative values indicate systemic failure) |
| Conditional Loss | Selection effect on payoffs |
| Incoherence | Variance-to-error ratio across simulation replays |
| Illusion Delta | Gap between perceived coherence and actual consistency |
Scenario files define agent families (e.g., LDT, RLM, Scholar), governance levers, and specific success thresholds for toxicity and welfare.
SWARM Safety Advanced Features
- Support for 38 distinct agent implementations, including Reinforcement Learning from Memory (RLM) and Council-based deliberation.
- 29 modular governance levers such as Reputation Decay, Sybil Detection, and Incoherence Friction.
- Multi-framework bridges for Openclaw Skills, Claude Code, and DeepMind's Concordia.
- Complex phase transition analysis to identify systemic tipping points in agent swarms.
- Local-first API for secure agent registration, scenario submission, and real-time simulation management.
SKILL.md
Loading
Related Openclaw Skills
justinx
A real-time data streaming connector that pipes MQTT, Kafka, and webhook data directly into AI agents via the Model Context Protocol.
Idea Clawdbot
Launch autonomous background sessions to perform deep-dive market research and technical analysis on any business idea.
SearXNG Bangs
A privacy-respecting metasearch skill that aggregates results from over 250 engines with support for DuckDuckGo-style bangs.
Daily Company Briefing
An automated business intelligence agent that aggregates corporate news, partner updates, and industry bidding opportunities into a concise morning briefing.
SWARM Safety
A sophisticated research framework for studying emergent risks and governance paradoxes in multi-agent AI systems through probabilistic assessment.
Boktoshi Human My-Endpoints Helper
A specialized helper skill for accessing authenticated human account endpoints on the Boktoshi platform via Firebase ID tokens.
METADATA
Featured*
