AI Agent Swarm¶
The COOWN AI swarm is a team of 8 specialized AI agents that support business operations — from strategic planning and compliance to development, design, and quality assurance. The swarm is orchestrated by Maya (🐝), the default agent and team leader.
What It Is¶
A multi-agent system where each agent has a defined role, personality (SOUL.md), memory, and access controls. Agents collaborate through task delegation, and all operate under human oversight. The swarm handles documentation, compliance checks, code review, design feedback, infrastructure monitoring, and quality assurance — augmenting human capacity rather than replacing it.
Tech Stack¶
| Layer | Technology | Purpose |
|---|---|---|
| Orchestration | OpenClaw v2026.2.x | Multi-agent management, session routing, memory, tool policies |
| Cloud models | Anthropic Claude (Opus 4.6, Sonnet 4.5, Haiku 3.5) | Primary reasoning for most agents |
| Cloud models | OpenAI (GPT-4o, GPT-4o Mini, o1) | Fallback and specialized tasks |
| Local models | Ollama (Llama 3.2, Qwen3 8B, Qwen3 Coder 30B) | Budget agents, fallback, confidential data |
| Image generation | Gemini (via nano-banana-pro skill) | Design agent (Leonardo) |
| Version control | Git + GitHub (coown-box org) |
Knowledge base (Tech-Pub), operations (Dev-Ops) |
| Hardware | Dedicated Linux workstation | On-premise, firewall-isolated |
| Google Workspace | gog CLI via service account + domain-wide delegation |
Gmail, Calendar, Drive access |
| Messaging | Telegram (OpenClaw channel, DM-only, allowlisted) | Real-time communication with CEO |
| Communication | Dedicated email (maya@coown.foundation), Git account (Maya-COOWN) | Isolated from production systems |
Infrastructure¶
Hardware¶
| Component | Specification |
|---|---|
| Machine | ASUS TUF Z370-PRO GAMING |
| CPU | Intel Core i7-8700K × 12 |
| RAM | 40 GB |
| GPU | NVIDIA GeForce GTX 1070 Ti |
| Storage | 1 TB SSD (encrypted) |
| OS | Ubuntu 25.04 (64-bit, Wayland, headless server mode) |
Network & Security¶
The AI swarm runs on a firewall-separated "dirty server" in Simon's home office:
- No write access to critical production data
- No read access to confidential customer data
- Network-isolated from production infrastructure
- Encrypted disk (LUKS) with keyring-protected boot
- Ollama binds to 127.0.0.1 only (no network exposure)
API Accounts & Credentials¶
All API keys are created by Simon using the info@managed.com identity:
| Service | URL | Purpose |
|---|---|---|
| Anthropic | platform.claude.com | Claude Opus, Sonnet, Haiku API keys |
| OpenAI | platform.openai.com/api-keys | GPT-4o, o1 API keys |
| Google AI Studio | aistudio.google.com/app/api-keys | Gemini API key (image generation) |
| GitHub | github.com (Maya-COOWN account) |
Knowledge base repos (Tech-Pub, Dev-Ops) |
API keys are stored in per-agent auth-profiles.json files under ~/.openclaw/agents/<Name>/agent/. Agents never see raw keys — OpenClaw injects them at runtime.
Configuration Principles¶
- Single config file — All agent definitions, models, routing, and skills are declared in one file:
openclaw.json - Model failover — Every agent has a primary model and fallbacks. Cloud fails → local model takes over automatically (zero downtime)
- Workspace isolation — Each agent has its own workspace with SOUL.md, memory, and tools. Maya (default) shares the main workspace with Tech-Pub and Dev-Ops repos
- Least privilege — Agents get minimum permissions for their role. Only Maya can spawn sub-agents to all other agents
- Human checkpoints — No autonomous external communications. Strategic decisions escalate to Simon (CEO)
- Local-first for sensitive data — Confidential data is processed exclusively on local models (Ollama). Cloud models only see public/internal data
Live Configuration¶
The swarm runs on the dedicated Linux workstation described above. The live configuration is at:
/home/skai8888/.openclaw/
├── openclaw.json # Master config (agents, models, skills, gateway)
├── workspace/ # Maya's workspace (default agent)
│ ├── SOUL.md # Maya's personality and rules
│ ├── Tech-Pub/ # Knowledge base repo (Git)
│ ├── Dev-Ops/ # Operations repo (Git)
│ └── memory/ # Daily logs
└── agents/ # Specialist agent workspaces
├── General/workspace/ # Non-COOWN tasks
├── Saul/workspace/
├── Rex/workspace/
├── Paul/workspace/
├── Leonardo/workspace/
├── Devin/workspace/
└── Tom/workspace/
Agent Roster¶
| Agent | ID | Role | Primary Model | Fallbacks | Status |
|---|---|---|---|---|---|
| 🐝 Maya | main |
Manager & Coordinator | Claude Opus 4.6 | Sonnet 4.5 → Llama 3.2 | ✅ Active |
| 🌐 General | general-001 |
General Purpose (non-COOWN) | Claude Opus 4.6 | Sonnet 4.5 → Llama 3.2 | ✅ Active |
| 💼 Saul | commercial-001 |
Commercial & Sales | Claude Opus 4.6 | Sonnet 4.5 → Llama 3.2 | ✅ Active |
| ⚖️ Rex | compliance-001 |
Compliance & Legal | Llama 3.2 (local) | Opus 4.6 → Qwen3 8B | ✅ Active |
| 🔧 Paul | infra-001 |
Infrastructure & DevOps | Claude Haiku 3.5 | Llama 3.2 | ✅ Active |
| 🎨 Leonardo | design-001 |
Design & Creative | Claude Sonnet 4.5 | GPT-4o → Llama 3.2 | ✅ Active |
| 💻 Devin | dev-001 |
Development & Architecture | Claude Opus 4.6 | Qwen3 Coder 30B | ✅ Active |
| 🧪 Tom | testing-001 |
Testing & QA | Claude Opus 4.6 | Qwen3 Coder 30B | ✅ Active |
Agent profiles and workspace data: Dev-Ops: Agent Team Directory
Governance, access controls, and data handling: Agent Governance Policy
How to Work with Agents¶
For practical instructions on interacting with agents: → User Guide
Backups & Recovery¶
Configuration and workspace backups are stored in:
- Dev-Ops/archive/ai-swarm-backups/ — tactical archive on GitHub (config snapshots, SOUL backups)
- ~/archive/baselines/YYYY-Wnn/ — weekly full baselines (local, offered to Simon for Proton Drive transfer)
- ~/archive/openclaw/config-backups/ — config snapshots on change (30-day local retention)
Git history provides automatic version tracking for all committed content.
Recovery procedures: → Backup & Recovery Plan → WI-003 Backup & Recovery
Data Handling & Security¶
| Sensitivity | Allowed Models | Examples |
|---|---|---|
| Public | Any (cloud or local) | Published docs, open-source |
| Internal | Cloud with caution, prefer local | Project plans, process docs |
| Confidential | Local only (Ollama) | Customer data, financial info |
| Secret | No agent access | Credentials, legal contracts |
Full policy: Agent Governance Policy §8
Knowledge Base¶
Agents access IMS documentation via file reads and OpenClaw's built-in memory_search (semantic search over workspace files).
Scope: Tech-Pub docs/ content (excluding archive/ and logs)
Not indexed: Dev-Ops, ~/archive/, backup files
Maintenance¶
| Task | Owner | Frequency |
|---|---|---|
| Documentation sync (GitHub ↔ workspace) | Maya (cron) | Every 4 hours |
| Weekly baseline archive | Paul (cron) | Weekly (Monday 07:00) |
| Session log review & incidents | Paul (cron) | Daily (06:00) |
| Memory pruning | Maya (cron) | Monthly (1st, 09:00) |
| Link & redundancy checks | Tom (cron) | Monthly (15th, 09:00) |
| PDCA management review | Maya + Simon | Quarterly |
Retention details: Data Retention Policy §7
Related Documents¶
- Agent Governance Policy
- Data Retention Policy
- Backup & Recovery Plan
- User Guide
- WI-001 File Management
- WI-002 Archival Process
- WI-003 Backup & Recovery
- Agent Team Directory (Dev-Ops)
Last Updated: 2026-02-10
Maintainer: Paul