Building a Multi-Agent AI with LangGraph, Gemini Pro, and Real-Time Tools

Introduction

AI is evolving from simple question-answer systems to autonomous multi-agent architectures. Instead of a single model trying to handle everything, multi-agent AI combines specialized agents—each expert in a domain (like weather, finance, or research)—with an orchestrator that decides which agent to invoke.

In this tutorial, we’ll explore how to build a production-ready multi-agent AI using:

• LangGraph for orchestration
• Gemini Pro 2.5 (Google Generative AI) for fallback reasoning
• Real-time APIs (weather, currency, stocks, Wikipedia)
• Math.js for safe mathematical computation

Why Multi-Agent AI?

Unlike a single monolithic model, multi-agent systems have several advantages:

• Domain Expertise: Each agent specializes (e.g., WeatherAgent, FinanceAgent).
• Improved Accuracy: Delegates tasks to the most relevant tool or API.
• Extensibility: Easily add new tools like News API, LangChain memory, or Telegram bot integration.
• Cost Efficiency: Only call LLMs when other tools cannot answer.

Scope of This Project

Our multi-agent AI includes:

• WeatherAgent: Fetches live weather data (OpenWeatherMap API).
• MathAgent: Evaluates secure mathematical expressions using math.js.
• FinanceAgent: Currency conversion & stock price lookup.
• ResearchAgent: Wikipedia summaries (and extendable to News APIs).
• Orchestrator: Routes queries intelligently using LangGraph & Gemini fallback.

The architecture is flexible and modular—you can add memory (vector DB), integrate LangChain, or deploy it as a Telegram bot or web app.

How It Works

1. Input Query: User asks a question like:
   “Convert 125 USD to INR”
   “What’s the weather in London?”
   “Wiki Alan Turing”
2. Router Agent: Uses LangGraph’s state machine to pick the right agent.
3. Specialized Agents:
   • Weather API → OpenWeatherMap
   • Currency → Frankfurter API (no key)
   • Stocks → Alpha Vantage API
   • Wikipedia → REST summary API
   • Math → math.js
4. Fallback: If no specialized agent matches, query is sent to Gemini Pro for a natural language answer.

Implementation: Step by Step

We implemented the solution in Node.js (ESM) using:

• dotenv for API keys
• axios for HTTP calls
• mathjs for safe math evaluation
• LangGraph for agent orchestration
• @google/generative-ai for LLM responses

Key Features:

• Regex-based query parsing for extracting city names, stock symbols, and currency pairs.
• Safe evaluation of math expressions (no eval!).
• Fallback to Gemini Pro for open-ended queries.
• Error handling & sanitization for production readiness.

Sample Queries & Outputs

• What is the weather in London? → WeatherAgent fetches from OpenWeatherMap.
• Convert 125 USD to INR → FinanceAgent calls Frankfurter API.
• Wiki Alan Turing → Wikipedia summary.
• Calculate (56*23)+100 → MathAgent computes result.
• Give me a motivational quote → Gemini fallback.

Git code is available at : https://github.com/110059/agentic-ai