AI Agent Orchestrator System

Personal Work Information Centralization Without Complex Integration

πŸ₯‡ 1st Place Winner Team

Team: λ„ˆλ„ ν• μˆ˜μžˆμ–΄ (You Can Do It)

H. Choi (Agent) β€’ S. Bae (기획자) β€’ J. Lee, J. Lee (MCP Server) β€’ J. Hwang (Backoffice)

September 30 - October 2, 2025 β€’ Meta x Llama Academy

Abstract

Existing methods for personal work information management typically require complex integrations and manual data entry across multiple AI tools, messengers, and document applications. This leads to scattered information, difficulty in priority management, and repetitive context explanations. In this work, we propose a simpler approach: we first collect all work-related conversations and data using the Model Context Protocol (MCP), then orchestrate intelligent agents using LangGraph to automatically process, summarize, and provide daily briefings. We demonstrate that this strategy achieves seamless integration across multiple AI clients (ChatGPT Desktop, Claude Desktop/CODE, Cursor) and productivity tools (Gmail, Slack, Notion) while maintaining 3D consistency in data flow and user experience.

How It Works

1. Data Collection

MCP protocol automatically captures conversations from AI clients and productivity tools into our Context Registry database.

2. Agent Orchestration

LangGraph StateGraph manages the workflow: plan β†’ cr_read β†’ summarize β†’ cr_write for intelligent processing.

3. Daily Briefing

Automated collection of Gmail, Slack, and Notion data, processed by AI models and delivered as structured Notion pages.

4. Backoffice Interface

Web-based management interface for viewing registry data, monitoring agent flows, and scheduling background jobs.

Key Features

  • Multi-AI Client Support: Seamless integration with ChatGPT Desktop, Claude Desktop/CODE, Cursor, and Notion
  • Real-time Conversation Logging: Automatic recording through standardized MCP communication
  • LangGraph-based Workflow: Intelligent agent orchestration with plan β†’ cr_read β†’ summarize β†’ cr_write process
  • Context Registry: Lightweight SQLite database for permanent storage and retrieval
  • Automated Daily Briefing: AI-powered summarization and Notion page generation at 07:00 KST daily
  • Backoffice Web Interface: Registry Viewer, Agent Flow Monitor, and Background Job scheduling

Technical Stack

Backend

Python 3.10+ FastAPI LangGraph SQLite MCP Protocol

AI & ML

Ollama + LLaMA 3.2 Upstage Solar OpenAI API

Frontend

Jinja2 HTML/CSS/JS Bulma CSS

DevOps & Tools

uv APScheduler aiohttp