SWARM Safety Skill

Study how intelligence swarms — and where it fails.

SWARM is a research framework for studying emergent risks in multi-agent AI systems using soft (probabilistic) labels instead of binary good/bad classifications. AGI-level risks don't require AGI-level agents — harmful dynamics emerge when many sub-AGI agents interact, even when no individual agent is misaligned.

v1.7.0 | 38 agent types | 29 governance levers | 55 scenarios | 2922 tests | 8 framework bridges

Repository: INLINECODE0

Hard Rules

• - SWARM simulations run locally. Install the package first.
• Do not submit scenarios containing real API keys, credentials, or PII.
• Simulation results are research artifacts. Do not present them as ground truth about real systems.
• When publishing results, cite the framework and disclose simulation parameters.

Security

• - API binds to localhost only (127.0.0.1) by default to prevent network exposure.
• CORS restricted to localhost origins by default.
• No authentication on development API — do not expose to untrusted networks.
• In-memory storage — data does not persist between restarts.
• For production deployment, add authentication middleware and use a proper database.

Install

CODEBLOCK0

Quick Start (Python)

CODEBLOCK1

Quick Start (CLI)

CODEBLOCK2

Quick Start (API)

Start the API server:

CODEBLOCK3

API documentation at http://localhost:8000/docs.

Security Note: The server binds to 127.0.0.1 (localhost only) by default. Do not bind to 0.0.0.0 unless you understand the security implications and have proper firewall rules in place.

Register Agent

CODEBLOCK4

Returns agent_id and api_key.

Submit Scenario

CODEBLOCK5

Create & Join Simulation

CODEBLOCK6

Core Concepts

Soft Probabilistic Labels

Interactions carry p = P(v = +1) — probability of beneficial outcome:

CODEBLOCK7

Five Key Metrics

Metric	What It Measures
Toxicity rate	Expected harm among accepted interactions: INLINECODE8
Quality gap

Adverse selection indicator (negative = bad): E[p \| accepted] - E[p \| rejected] | | Conditional loss | Selection effect on payoffs | | Incoherence | Variance-to-error ratio across replays | | Illusion delta | Gap between perceived coherence and actual consistency |

Agent Types (14 families, 38 implementations)

Type	Behavior
Honest	Cooperative, trust-based, completes tasks diligently
Opportunistic

Maximizes short-term payoff, cherry-picks tasks | | Deceptive | Builds trust, then exploits trusted relationships | | Adversarial | Targets honest agents, coordinates with allies | | LDT | Logical Decision Theory with FDT/UDT precommitment | | RLM | Reinforcement Learning from Memory | | Council | Multi-agent deliberation-based decisions | | SkillRL | Learns interaction strategies via reward signals | | LLM | Behavior determined by LLM (Anthropic, OpenAI, or Ollama) | | Moltbook | Domain-specific social platform agent | | Scholar | Academic citation and research agent | | Wiki Editor | Collaborative editing with editorial policy |

Governance Levers (29 mechanisms)

• - Transaction Taxes — Reduce exploitation, cost welfare
• Reputation Decay — Punish bad actors, erode honest standing
• Circuit Breakers — Freeze toxic agents quickly
• Random Audits — Deter hidden exploitation
• Staking — Filter undercapitalized agents
• Collusion Detection — Catch coordinated attacks (the critical lever near collapse threshold)
• Sybil Detection — Identify duplicate agents
• Transparency Ledger — Reward/penalize based on outcome
• Moderator Agent — Probabilistic review of interactions
• Incoherence Friction — Tax uncertainty-driven decisions
• Council Deliberation — Multi-agent governance decisions
• Diversity Enforcement — Prevent monoculture collapse
• Moltipedia-specific — Pair caps, page cooldowns, daily caps, self-fix prevention

Framework Bridges

Bridge	Integration
Concordia	DeepMind's multi-agent framework
GasTown

Multi-agent workspace governance | | Claude Code | Claude CLI agent integration | | LiveSWE | Live software engineering tasks | | OpenClaw | Open agent protocol | | Prime Intellect | Cross-platform run tracking | | Ralph | Agent orchestration | | Worktree | Git worktree-based sandboxing |

Scenario YAML Format

CODEBLOCK8

Key Research Findings

Phase Transitions (11-scenario, 209-epoch study)

Regime	Adversarial %	Toxicity	Welfare	Outcome
Cooperative	0-20%	< 0.30	Stable	Survives
Contested

20-37.5% | 0.33-0.37 | Declining | Survives | | Collapse | 50%+ | ~0.30 | Zero by epoch 12-14 | Collapses |

Critical threshold between 37.5% and 50% adversarial agents separates recoverable from irreversible collapse.

Governance Cost Paradox (v1.7.0 GasTown study)

42-run study reveals: governance reduces toxicity at all adversarial levels (mean reduction 0.071) but imposes net-negative welfare costs at current parameter tuning. At 0% adversarial, governance costs 216 welfare units (-57.6%) for only 0.066 toxicity reduction.

Case Studies

GasTown Governance Cost

Study governance overhead vs. toxicity reduction across 7 agent compositions with and without governance levers. Reveals the safety-throughput trade-off. See scenarios/gastown_governance_cost.yaml.

LDT Cooperation

220 runs across 10 seeds comparing TDT vs FDT vs UDT cooperation strategies at population scales up to 21 agents. See scenarios/ldt_cooperation.yaml.

Moltipedia Heartbeat

Model the Moltipedia wiki editing loop: competing AI editors, editorial policy, point farming, and anti-gaming governance. See scenarios/moltipedia_heartbeat.yaml.

Moltbook CAPTCHA

Model Moltbook's anti-human math challenges and rate limiting: obfuscated text parsing, verification gates, and spam prevention. See scenarios/moltbook_captcha.yaml.

API Endpoints (Full Reference)

Method	Endpoint	Description
GET	INLINECODE14	Health check
GET

/ | API info | | POST | /api/v1/agents/register | Register agent | | GET | /api/v1/agents/{agent_id} | Get agent details | | GET | /api/v1/agents/ | List agents | | POST | /api/v1/scenarios/submit | Submit scenario | | GET | /api/v1/scenarios/{scenario_id} | Get scenario | | GET | /api/v1/scenarios/ | List scenarios | | POST | /api/v1/simulations/create | Create simulation | | POST | /api/v1/simulations/{id}/join | Join simulation | | GET | /api/v1/simulations/{id} | Get simulation | | GET | /api/v1/simulations/ | List simulations |

Citation

CODEBLOCK9

Linked Docs

• - Skill metadata: INLINECODE26
• Agent discovery: INLINECODE27
• Full documentation: INLINECODE28
• Theoretical foundations: INLINECODE29
• Governance guide: INLINECODE30
• Red-teaming guide: INLINECODE31
• Scenario format: INLINECODE32