Inspiration
This summer, I went on a road trip to Yosemite. Planning it turned out to be much harder than I expected finding good spots, checking availability, building an itinerary, and coordinating everything was a huge task.
Out of curiosity, I tried using ChatGPT to generate a travel itinerary. It looked great on paper until we got there and discovered that some of the places didn’t exist, others were closed for the season, and a few were listed under completely wrong names.
That’s when I realized the core problem: while AI can generate itineraries, they’re often unreliable and disconnected from real-world data. I wanted to fix that by building a travel planner that combined AI reasoning with structured collaboration between specialized agents.
What it does
TravelAdvisor is an AI-powered, multi-agent travel planning system that can generate a complete, executable trip plan from a single prompt.
You can ask something like:
“Plan a 5-day food and culture trip to Kyoto in April.”
TravelAdvisor then creates a team of specialized AI agents: a Planner, a Researcher, a Weather Analyst, Coordinator, each responsible for a unique part of the process.
They collaborate in a simulated “group chat”, exchanging ideas and refining details until they agree on a final, realistic itinerary. The Coordinator compiles everything into a clean, actionable travel plan and checks with the user for feedback or refinement before finalizing.
How we built it
The entire project was built in Python using Microsoft’s AutoGen framework, integrated with Gemini models for language understanding.
Each agent is programmed with a specific role and toolset. They communicate using a defined protocol that mimics a collaborative chatroom, allowing context sharing and problem-solving as a team instead of a single model doing everything.
The GroupChatManager coordinates turn-taking and ensures coherent conversation flow. The Coordinator agent merges the outputs from all other agents into one unified plan that is contextually accurate and realistic.
Challenges we ran into
The biggest challenge was framework instability and library conflicts. AutoGen’s structure is powerful but quite rigid in certain setups, especially when adapting it for non-OpenAI models like Gemini.
At one point, the entire system broke due to dependency mismatches and async handling errors. After extensive troubleshooting, I decided to switch frameworks last-minute, refactoring the code to maintain multi-agent functionality while ensuring the system still worked smoothly under time constraints.
Balancing collaborative agent design with reliable API calls and conversation flow was also a tough problem to solve.
Accomplishments that we're proud of
Successfully built a multi-agent collaboration system that can plan trips autonomously.
Created a seamless “group chat” protocol where AI agents discuss, refine, and reason together.
Integrated Gemini AI within an AutoGen-like framework despite compatibility challenges.
Designed a structure that can scale easily, new agents (like lodging experts or budget advisors) can be added effortlessly.
But most importantly, I’m proud that the system can produce accurate, reliable itineraries that actually make sense in the real world.
What we learned
Working with distributed AI agents requires a deep understanding of asynchronous reasoning, message passing, and context persistence.
Frameworks like AutoGen are incredibly powerful but can be brittle when adapting them beyond their default environment.
AI is most effective when multiple agents specialize and collaborate, rather than one large model trying to do everything.
Reliability in AI-generated content often depends on real-world verification and cross-agent reasoning.
What's next for TravelAdvisor
Build a web app interface using Flask or Next.js to make it accessible to anyone planning a trip.
Experiment with voice-based interaction and local caching for offline planning during travel.
Log in or sign up for Devpost to join the conversation.