MundoAgent

"Beyond chat: An autonomous Gemini-powered agent automating 2026 World Cup logistics and travel workflows via MongoDB MCP."

Comment

Inspiration

The 2026 FIFA World Cup is a historic event spanning three nations (USA, Mexico, and Canada) and 16 host cities[cite: 1]. While fans are thrilled, they are also facing an absolute logistical nightmare: cross-border travels, volatile match schedules, hyper-inflated hotel clusters, and complex local transit grids[cite: 1].

As text-based AI chatbots only provide static, generic answers, they leave the user to manually handle all the heavy lifting of multi-step planning and itinerary execution[cite: 1]. I wanted to build a solution that moves entirely beyond the chat interface. Inspired by autonomous enterprise agents, I developed MundoAgent to act as a proactive, reliable digital concierge that transforms chaotic World Cup logistics into a fluid, automated workflow under human oversight[cite: 1].

What it does

MundoAgent is a fully functional AI agent designed to orchestrate and automate the travel lifecycle for 2026 World Cup attendees. Instead of answering queries with text walls, the agent performs active context-driven reasoning and execution phases:

  • Dynamic Database Querying: It connects via Model Context Protocol (MCP) to analyze dynamic match schedules, live stadium venues, and adjacent transport systems directly within a MongoDB environment[cite: 1].
  • Intelligent Accommodation Scoring: It processes local hotel metrics and computes optimal staging areas based on budget and transit weights using real-time heuristics[cite: 1].
  • Stateful Automated Booking: Once the multi-step trip itinerary is verified and explicitly approved by the fan, MundoAgent bypasses manual input and writes the stateful transaction data directly back to the database backend[cite: 1].

How we built it

The core infrastructure of MundoAgent is highly decoupled and engineered using industry-grade tools for modern AI environments:

  1. The Core Orchestration Engine: Built using Google Cloud Agent Builder driven by Gemini 3[cite: 1]. This runtime layer hosts the core reasoning loops, conversational state handlers, and custom tool parameters[cite: 1].
  2. The Communications Standard (MCP): Integrated the open-source Model Context Protocol (MCP) via a custom Node.js server[cite: 1]. Utilizing stdio transport abstractions, it registers precise tool schemas that expose programmatic capabilities directly into Gemini’s context window[cite: 1].
  3. The Persistent Data Layer: Powered by MongoDB Atlas, structured into document collections for match/stadium realities (stadiums_and_matches) and active user-approved workflows (fan_itineraries)[cite: 1].

Challenges we ran into

Transitioning from non-deterministic language models to deterministic, reliable software actions posed significant engineering constraints:

  • Mitigating Multi-Step Derailment: Complex travel loops can cause an LLM to stray from sequential dependency pathways. I implemented rigorous system prompt framing and structured execution steps within Google Cloud Agent Builder to guarantee strict logical guardrails[cite: 1].
  • Schema Mapping over Stdio Channels: Structuring dynamic language statements into pristine JSON objects for safe database insertion required precise schema validation rules on the Node.js MCP server[cite: 1].
  • Mathematical Weighting for Logistics Optimization: Evaluating accommodation sectors dynamically required a rigid formulation. Let $P$ define the average nightly rate in a zone, and $D$ define the stadium proximity vector in kilometers[cite: 1]. To solve the non-linear choice conflict for the agent, I implemented a customized scaling heuristic[cite: 1]:

$$\min_{zone} \quad S = w_1 \left( \frac{P}{P_{max}} \right) + w_2 \left( \frac{D}{D_{max}} \right)$$

Where $w_1$ and $w_2$ are weight attributes modulated dynamically by the agent based on user-stated constraints (such as strict budget compliance versus rapid walking proximity)[cite: 1].

Accomplishments that we're proud of

  • True Autonomous Behavior: Successfully broke out of the generic "read-only" chatbot ecosystem into an agent that plans and modifies real operational databases under continuous human control[cite: 1].
  • Production-Grade MCP Integration: Built a fully operational, stable Model Context Protocol pipeline with MongoDB from scratch, allowing a large language model to safely interact with local runtime infrastructure[cite: 1].
  • Robust Logic & Analytical Framework: Blended macroeconomic and analytic perspectives with AI engineering to build an environment that solves actual, concrete logistical inflation challenges for real users[cite: 1].

What we learned

This hackathon provided an exceptional learning curve regarding stateful AI systems[cite: 1]. I gained a profound understanding of the Model Context Protocol, seeing firsthand how it unifies how models ingest outside tools and databases without relying on fragile custom API layers[cite: 1]. I also mastered the utilization of Google Cloud Agent Builder to coordinate complex prompt guidelines, and learned how a structured database like MongoDB acts as an elegant memory mesh that transforms ephemeral AI conversations into reliable, production-ready enterprise transactions[cite: 1].

What's next for MundoAgent

  • Live Multi-Partner Scalability: Integrating Fivetran data pipelines to fetch real-time airline flight prices and combining it with Elastic for hyper-fast semantic search over localized tourism data.
  • Cross-Border Compliance Agents: Incorporating legal check logic to auto-verify temporary visa requirements or transit access protocols across the US-Mexico-Canada borders during the World Cup.
  • Real-time Push Notifications: Bridging the database engine with live automated alert systems so the agent can rewrite itineraries instantly on behalf of fans if match schedules or local transport routes shift unexpectedly.

Built With

Share this project:

Updates