A multi-agent AI marketing system that coordinates 32 specialized agents to create, evaluate, and publish content across platforms, then learns from what works.
This is the Observatory, the real-time monitoring dashboard for Holus. Holus is a federated AI system that acts as an autonomous marketing strategist for a portfolio of AI products. It does not just generate content. It observes platform analytics, reasons about strategy using Claude Opus, dispatches work to specialized agents, evaluates every output with domain-expert judges, and feeds results back into the next cycle. The Observatory shows this entire loop in real time.
The system uses a federated architecture: Holus holds the brain (strategy, decisions, learning) while independent silo services handle execution (video editing, image generation, publishing). Communication happens via MCP (Model Context Protocol) tool calls, not shared databases.
Reads analytics from social media platforms. What performed well? What audience segments are growing? Which content pillars convert?
Claude Opus analyzes patterns across 30 days of data. Decides what content to create, for which product, on which platform.
Dispatches to specialized agents: hook writers, blog writers, carousel architects, SEO researchers. Each agent has domain-specific expertise.
7 domain-expert judges score every piece. Written content, visual content, and brand safety each have dedicated evaluators with custom rubrics.
Strategy and self-improvement orchestration
6 content categories: authority, visual, video, growth, research, repurposing
Domain-expert quality judges with category-specific rubrics
Code quality, security auditing, knowledge management
Orchestration
Python + LangGraph
Strategy Model
Claude Opus 4.6
Content Models
Claude Sonnet 4.6
Research
Gemini 2.5 Pro
Event Bus
Redis pub/sub
Silo Protocol
MCP (Model Context Protocol)
Observatory
Next.js 16 + Tailwind
Observatory API
FastAPI
State
JSONL + YAML (no database)