🧠 Enterprise Multi-Agent AI Engine

An advanced, highly scalable Multi-Agent AI platform built for enterprise microservice orchestration. This system leverages LangGraph for stateful agent workflows and the Model Context Protocol (MCP) to dynamically decouple AI capabilities from backend executions.

🌟 Overview

This project serves as the intelligent core of an Enterprise App Builder Studio. Instead of relying on monolithic LLM prompts, it utilizes a Domain-Driven Design (DDD) approach where autonomous AI sub-agents (Orchestrators and Executors) collaborate to interact with users, validate complex inputs (like state-management code), and securely execute database mutations across a Kubernetes-based microservice cluster.

🏗️ Core Architecture

The system is built on an Orchestrator-Executor Pattern:

• AiEngine (The Brain): A FastAPI-based microservice that manages chat threads, session memory, and LLM inferences.
• Orchestrator Agents: Act as the user-facing interface. They are responsible for understanding intent, maintaining scope boundaries (e.g., refusing UI design tasks if they are Data agents), and performing interactive slot-filling.
• Executor Agents (Compiled Sub-Agents): Specialized workers that receive validated tasks from the Orchestrator and interact strictly with backend databases via MCP tools.
• Consul Service Discovery: AI tools dynamically resolve target internal APIs (like app-store-server) using Consul, ensuring smooth operation in distributed environments.

✨ Key Features & AI Capabilities

• 🔌 Model Context Protocol (MCP) Integration: Tools are exposed via unified MCP servers (AppBuilderTools). The AI engine consumes these tools dynamically without needing hardcoded API routes.
• 🛡️ Infrastructure & Network Guards: Built-in prompt engineering to detect underlying network timeouts (e.g., TaskGroup cluster routing failures). The AI gracefully halts execution and reports the outage instead of falling into infinite retry loops, drastically saving server resources.
• 📋 Interactive Slot Filling: The AI does not blindly fail on missing parameters. It initiates a step-by-step interview process to gather required fields before executing database mutations.
• 🔬 Smart Code Validation: The Orchestrator parses user-provided code (e.g., Zustand state logic in JavaScript). It actively counts parameters inside the raw code and blocks execution if it contradicts the user's claims, preventing bad data from reaching the database.
• 🚧 Strict Scope Boundaries: Prompts are engineered to isolate domains. The Public Data agent will cleanly refuse tasks meant for the Page Builder agent and suggest proactive alternatives.

💻 Tech Stack

• Backend Framework: FastAPI, Uvicorn, Python 3.10+
• AI & LLM Orchestration: LangGraph, LangChain
• Tooling Protocol: Model Context Protocol (MCP)
• Service Discovery: HashiCorp Consul
• Database: PostgreSQL (Internal Cluster)
• LLM Provider: Minimax / Qwen (via Custom Gateway)