Multi-Agent System for Travel Planning

A multi-agent travel planning system demonstrating the SubAgents pattern from LangChain/LangGraph. A central Supervisor Agent coordinates four specialized subagents to produce comprehensive, end-to-end travel plans from a single natural language request.

Architecture

The system follows a hub-and-spoke pattern: the Supervisor Agent receives the user's request, decides which specialists to invoke, and synthesizes their results into a cohesive response.

• Supervisor Agent — routes requests and assembles the final travel plan
• Flights Agent — searches and compares flight options by destination, budget, and stop preference
• Hotels Agent — searches accommodations matching location and budget constraints
• Activities Agent — recommends things to do and places to eat at the destination
• Itinerary Agent — composes a day-by-day travel schedule from the other agents' outputs

Key Features

• SubAgents pattern — each subagent is wrapped as a @tool and exposed to the Supervisor via LangChain's tool-calling interface, keeping concerns cleanly separated
• Conversation memory — InMemorySaver checkpointer enables multi-turn follow-up questions scoped by thread_id
• Composable agents — new specialist agents can be added by creating a subagent file and registering its wrapper tool with the Supervisor
• Streaming output — responses are streamed via agent.stream() for real-time feedback
• Mock data layer — a tools/mock_data.py module provides realistic travel data, designed to be swapped for real APIs (Amadeus, Booking.com, TripAdvisor, etc.)

Tech Stack

• LangGraph / LangChain — agent orchestration, tool node, create_agent API
• OpenAI / Anthropic — language model backends (configurable)
• Python — implementation language