Guardian System Architecture

Overview

Guardian is a production multi-agent AI system running 49 specialized agents 24/7. This document describes the full architecture.

System Layers

Layer 4: User Interface
  - Slack (#elliott-codex) for delegation and reporting
  - VRChat chatbox (OSC protocol) for real-time interaction
  - Telegram bot for mobile alerts
  - CLI API (REST) for direct agent invocation

Layer 3: Orchestration
  - Orchestrator (guardian-orchestrateur): central task router and conflict resolver
  - Setup Master: pre-task blueprint generator
  - Brain V3: DeepSeek-powered supervisor for analytical routing
  - Task Registry: SQLite anti-duplication coordinator

Layer 2: Specialist Agents (49 total)
  - Revenue (5): Business, Comet, Arena, Money, Revenue Ops
  - Security (4): HAK, Ghost, Miracle Guard, GSIS Sentinel
  - Trading (3): Trading ICT, Signals, Market Analyzer
  - Infrastructure (6): Cloudflare, N8N, Docker, DNS, Updater, Watchdog
  - Intelligence (4): Brain, Deep Research, Architect, MCP Builder
  - Creative (5): Voice, VRChat Persona, Avatar, Charme, Translator
  - Utility (19): Codex, Memory, Pulse, Nexus, and more
  - Orchestration (3): Orchestrateur, Elliott (persona), Setup Master

Layer 1: Infrastructure
  - N8N VPS (65 workflows, 51 active)
  - Docker (8 containers: ChromaDB, Redis, N8N, LLM services)
  - Knowledge Graph (787 entities, 943 relations)
  - Cloudflare Tunnel (6 services exposed securely)
  - Local machine (RTX 5080, Windows)

Communication Patterns

Agent-to-Agent

• Slack channels: Formal delegation and reporting
• Task Registry: Anti-duplication coordination (SQLite)
• Agent Inbox: File-based message queue (26 agents)
• Knowledge Graph: Shared persistent memory (MCP)

Agent-to-Human

• VRChat chatbox: Real-time voice/text interaction via OSC
• Telegram: Mobile push notifications and commands
• Slack: Asynchronous reporting and delegation

Agent-to-External

• N8N webhooks: 14 operational webhook endpoints
• CLI API: REST interface for programmatic access
• Gmail MCP: Direct email drafting and sending
• LinkedIn MCP: Social media posting

Data Flow

Task Input (Slack/VRChat/Telegram/Direct)
    |
    v
Orchestrator (classify, blueprint, deduplicate)
    |
    v
Setup Master (identify agents, tools, risks)
    |
    v
Task Registry (claim, anti-duplicate check)
    |
    v
Specialist Agent(s) (execute with tools)
    |
    v
Quality Gate (verify output meets criteria)
    |
    v
Progress Update (progress.json + Slack report)

Production Stats

Metric	Value
Tasks processed	10,000+
Autonomous sessions	200+
Knowledge graph entities	787
N8N workflows	65 (51 active)
Webhooks	14/14 operational
Agent prompts	49 in production
Docker containers	8/8 running
Infrastructure cost	Under $50/month
Model cost	~$200/month (Claude Max)

Technology Stack

• LLM: Claude Opus/Sonnet 4.6 (primary), DeepSeek (N8N brains)
• Orchestration: N8N (workflow automation), custom task registry
• Memory: ChromaDB (vector), Knowledge Graph (MCP), file-based
• Communication: Slack API, Telegram Bot API, VRChat OSC
• Infrastructure: Cloudflare Tunnel, Docker, Hostinger VPS
• Security: Miracle Guard, GSIS Sentinel, credential rotation

---

This architecture has been running in production since January 2026.