Add SWARM to Multi-Agent Simulation Projects

[rsavitt]

Add SWARM to Multi-Agent Simulation Projects

Hi! I'd like to suggest adding SWARM to the Multi-Agent Simulation Projects section.

What is SWARM?

SWARM (System-Wide Assessment of Risk in Multi-agent systems) is a research framework for studying emergent risks and distributional safety in multi-agent AI systems. It focuses on how harmful dynamics can emerge from the interaction of many LLM-powered agents — even when no single agent is individually dangerous.

Key Features

• LLM Agent Support — First-class integration with Anthropic, OpenAI, and Ollama; configurable personas (Honest, Strategic, Adversarial); prompt audit logging and cost tracking.
• Probabilistic Safety Metrics — Soft (probabilistic) labels instead of binary classification: toxicity rate, quality gap (adverse selection detection), conditional loss, and incoherence.
• Configurable Governance Levers — Transaction taxes, reputation decay, circuit breakers, random audits, staking, collusion detection, sybil detection, and more.
• Multi-Agent Simulation — Heterogeneous agent populations (honest, opportunistic, deceptive, adversarial, LLM-backed) interacting in a marketplace environment.
• Red-Teaming Framework — Systematically test governance mechanisms against adaptive adversaries.
• Reproducible Research — YAML scenario definitions, seed-based reproducibility, append-only event logs, CSV/JSON exports.

Why it fits this list

SWARM directly addresses multi-agent cooperation and competition among LLM-powered agents, with a specific focus on safety and governance — a critical gap in the current ecosystem. It complements existing entries like AgentVerse and ChatArena by focusing on emergent risk measurement rather than general collaboration or game environments.

Entry

* ![SWARM Stars](https://img.shields.io/github/stars/swarm-ai-safety/swarm) [SWARM](https://github.com/swarm-ai-safety/swarm) - A research framework for studying emergent risks and governance in multi-agent LLM systems, featuring probabilistic safety metrics, configurable governance levers, and first-class LLM agent support.

Thank you for maintaining this excellent resource!