| Metric           | Before    | After   |
| ---------------- | --------- | ------- |
| Tracked sessions | 423/727   | 727/727 |
| Qdrant vectors   | 5,404     | 6,922   |
| Untracked gap    | 315 files | 0       |

- '/var/run/docker.sock:/var/run/docker.sock'

| Component          | Status                                                                                   |
| ------------------ | ---------------------------------------------------------------------------------------- |
| m2 (me)            | ✅ Fully operational — Telegram, memory (10K vectors), 17 skills, voice, browser, desktop |
| Planka             | ✅ Running, API works, boards + governance created                                        |
| Shared infra       | ✅ Qdrant, BGE-M3 embeddings, Speaches STT — all healthy on Coolify network               |
| miauczek container | 🟡 Running but unhealthy — no bot token, no verified API key                             |
| peter container    | 🟡 Running but unhealthy — no bot token, no API key                                      |
| pittbull container | 🟡 Running but unhealthy — no bot token, no API key                                      |

Peter: discovers market trend → signals "crypto shift"
Pittbull: notices API costs spiking → signals "infra costs"
Miauczek: finds regulatory news → signals "EU AI regulation"

Federation index: 3 cost-related signals in 48h
→ auto-convergence card created

Each agent decides independently: relevant to me?
If yes → they self-organize
If no → acknowledge, move on

fleet-governance/
├── CONSTITUTION.md ← Core principles (changes slowly)
├── VERSION ← 0.1.0
├── protocols/
│ ├── signals.md ← How agents talk to the field
│ ├── summits.md ← Strong alignment (rare)
│ ├── resources.md ← Shared services rules
│ └── onboarding.md ← New agent checklist
├── registry/
│ ├── agents.yaml ← Fleet roster + trust levels
│ └── infrastructure.yaml ← All shared services
├── alignment/
│ └── ALIGNMENT_MODEL.md ← The three-layer model
└── changelog/
└── CHANGELOG.md

./bin/compile workflow-templates/deploy-service.md
./output/deploy-service-skill/skill.sh --service_name test --git_repo https://github.com/test/repo
# ✅ All 5 steps executed successfully with full logging

cd ~/.openclaw/workspace/projects/dark-factory-engine
git remote add origin https://github.com/machine-machine/dark-factory-engine.git
git push -u origin master

cd ~/.openclaw/workspace/projects/dark-factory-engine  
./demo.sh

📁 dark-factory-engine/
├── 🔧 bin/compile           # Workflow → Skill compiler  
├── 📝 workflow-templates/   # Deploy Service + Run Tests examples
├── 🐍 compiler/            # Python parser + generator
├── 📊 output/              # Generated executable skills
└── 🎬 demo.sh              # End-to-end demonstration

git clone https://github.com/machine-machine/dark-factory-engine.git
cd dark-factory-engine
./demo.sh
# ✨ Compiles & executes both workflows successfully

// Instead of direct skill execution:
await executeSkill("deploy-service.sh", params);

// Use isolated subagent spawning:
await sessions_spawn({
  task: "Deploy service using our workflow",
  label: "deploy-user-api",
  cleanup: "keep", // or "delete" 
  runTimeoutSeconds: 600
});
// → Auto-announces completion back to requester

Dark Factory → Direct OpenClaw Skill → Execution

Dark Factory Workflow Engine
↓
sessions_spawn (isolated execution)
↓  
Subagent Registry (monitoring & control)
↓
Auto-announcement (completion notifications)

# workflow-templates/deploy-service.md  
name: "Deploy Service"
description: "Generic deployment workflow"
# No personality, inconsistent behavior

{
  "@type": "aieos:AIEntityObject",
  "identity": { "names": { "nickname": "Deployer" } },
  "psychology": {
    "neural_matrix": {
      "reliability": 0.95,  // Highly reliable
      "logic": 0.90,        // Very logical
      "creativity": 0.30    // Low creativity (follows procedures)
    }
  },
  "linguistics": {
    "text_style": {
      "style_descriptors": ["precise", "technical", "methodical"]
    }
  },
  "capabilities": {
    "skills": [
      { "name": "coolify-deployment", "priority": 1 },
      { "name": "health-monitoring", "priority": 2 }
    ]
  }
}

Workflow Template → Generic OpenClaw Skill → Execution

AIEOS Entity Definition
    ↓
Personality-Driven sessions_spawn
    ↓  
Consistent Specialist Behavior
    ↓
Auto-Announcement with Entity Voice
    ↓
Cross-Platform Entity Portability

// entities/devops-entity.json
{
  "neural_matrix": { "reliability": 0.95, "logic": 0.90 },
  "capabilities": {
    "skills": [
      { "name": "deploy-service", "priority": 1, "auto_activate": true }
    ]
  }
}

function compileAIEOSWorkflow(entity: AIEntityObject): OpenClawSkill {
  return {
    systemPrompt: buildPersonalityPrompt(entity.psychology),
    voiceConfig: entity.linguistics.voice.tts_config,
    skillPriorities: entity.capabilities.skills,
    behaviorWeights: entity.psychology.neural_matrix
  };
}

// Spawn entity-aware subagent
await sessions_spawn({
  task: "Deploy user API to staging",
  agentId: "devops-entity",  // Uses Deployer personality
  label: "staging-deployment"
});
// → Automatically applies DevOps personality, communication style, priorities

{
  "capabilities": {
    "skills": [
      { "name": "kubernetes-deployment", "priority": 1 },
      { "name": "terraform-provisioning", "priority": 2 },
      { "name": "monitoring-setup", "priority": 3 }
    ]
  }
}

interface ComplexAdaptiveSystem {
  self_organization: boolean;      // Agents form efficient structures spontaneously
  emergence: string[];            // System > sum of individual agents
  adaptation_rules: Rule[];       // Learning at individual + system levels
  feedback_loops: Loop[];         // Continuous improvement cycles
  network_effects: number;        // Value increases with connected agents
}

{
  "resource_allocation": "sealed_bid_auction",
  "coordination_protocol": "nash_equilibrium_seeking", 
  "incentive_alignment": "pareto_optimization",
  "conflict_resolution": "mechanism_design"
}

{
  "network_structure": "adaptive_small_world",
  "coordination_mechanism": "auction_based_resource_allocation", 
  "information_flow": "priority_weighted_broadcasting",
  "feedback_loops": ["performance", "learning", "adaptation"]
}

{
  "value_system": {
    "core_values": ["excellence", "collaboration", "innovation"],
    "decision_framework": "values_weighted_utility",
    "conflict_resolution": "restorative_justice"
  }
}

{
  "emergence_mechanisms": {
    "self_organization": true,
    "innovation_tournaments": "continuous", 
    "capability_evolution": "genetic_algorithm",
    "collective_intelligence": "swarm_optimization"
  }
}

# projects/ai-org-workforce/coordination.yaml
network_topology: "adaptive_small_world"
coordination_protocol: "yaml_driven_consensus" 
agents: [5 specialist entities with AIEOS personalities]
work_packages: [Phase 1 foundation, Phase 2 integration]
metrics: [individual KPIs, collective intelligence, emergence]

| Benchmark                     | What it tests                                               | Fit for AI Org                                                     |
| ----------------------------- | ----------------------------------------------------------- | ------------------------------------------------------------------ |
| AgentBench (Liu et al., 2023) | 8 environments: web, code, DB, game, OS tasks               | ✅ Good — tests real-world agent capability, measurable             |
| GAIA (Meta, 2023)             | Real-world questions requiring multi-step reasoning + tools | ✅ Strong — humans 92% vs GPT-4 15%; org should score higher        |
| SWE-bench                     | Real GitHub issues → code fixes                             | ✅ Perfect for your 5-agent org (different specialists per subtask) |
| HumanEval / MBPP              | Code generation accuracy                                    | ⚠️ Single-task, less interesting for org                           |
| WebArena                      | Web navigation + task completion                            | ⚠️ Less relevant to org intelligence                               |
| MMLU                          | Knowledge breadth                                           | ❌ Wrong layer — tests knowledge not coordination                   |

retro = cerebras_call(f"""
Analyze this AI org run. Score was {run_score}/100.
Output: {run_output}

Produce:
1. 5 memory entries (what each agent should remember)
2. 1 protocol suggestion (if score < 80 or improvement identified)
3. Role refinement notes (what each agent did well/poorly)
""")

# 2. Store memories
for memory in retro.memories:
qdrant.store(memory, tags={run_id, task_type, agent_id})

# 3. Protocol PR (if suggested)
if retro.protocol_suggestion:
git_pr(
repo="fleet-governance",
file="PROTOCOLS.md",
change=retro.protocol_suggestion,
auto_merge=(retro.confidence > 0.7)
)

# 4. Update learning log
append_to("LEARNING_LOG.md", {
run_id, task_type, score, delta_vs_prev, key_lessons
})

| Run   | Single | Multi-Org | Delta |        |
| ----- | ------ | --------- | ----- | ------ |
| Run 1 | 90     | 73        | −17   | SA won |
| Run 2 | 84     | 87        | +3    | MA won |
| Run 3 | 85     | 87        | +2    | MA won |

| Run | SA | MA | Delta |                           |
| --- | --- | --- | ----- | ------------------------- |
| 1   | 90 | 73 | −17   | baseline                  |
| 2   | 84 | 87 | +3    | structured JSON handoff ✅ |
| 3   | 85 | 87 | +2    | phase-locked synthesis ✅  |
| 4   | 92 | 68 | −24   | regression ❌              |

Never truncate specialist input. Full-context single synthesis call > split calls with partial context. If token limit hit: retry with +2000 tokens. If 3 attempts fail: use Claude for synthesis step only (synthesis is post-processing, not core agent execution).

| Run | SA | MA | Δ   |                                  |
| --- | --- | --- | --- | -------------------------------- |
| 1   | 90 | 73 | −17 | baseline                         |
| 2   | 84 | 87 | +3  | ✅ JSON handoff                   |
| 3   | 85 | 87 | +2  | ✅ phase structure                |
| 4   | 92 | 68 | −24 | ❌ split synthesis hallucinated   |
| 5   | 86 | 83 | −3  | ⚠️ synthesis fixed, domain drift |