Building MCP Before MCP Existed: Inside Despegar's Sofia Agent

2026-05-20 · Source: MLOps.community · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Intermediate, extended

Summary

Despegar, a Latin American travel company, developed Sofia, an AI concierge agent for travel, featuring a multi-agent architecture with a central "brain" called Chappie. Sofia handles various travel flows like flights, hotels, and activities, allowing different company squads to develop specialized agents. Initially built with a custom orchestration layer and a proprietary Multi-Agent Communication Protocol (MCP) before industry standards emerged, Sofia now integrates new MCPs for external connections. Available via WhatsApp and Despegar's app/website, Sofia assists with real-time offers, destination recommendations, and post-sales queries, achieving 30% client interaction. The agent covers the entire travel journey, from "dreaming" and planning to booking and in-trip support, aiming to perform complex tasks like flight changes and claim filing, offering capabilities beyond traditional platforms.

Key takeaway

For AI Product Managers developing customer-facing agents, recognize that users are willing to engage deeply for personalized travel experiences. Prioritize building a flexible, evolving multi-agent architecture that can integrate with diverse internal teams and external communication channels like WhatsApp. Focus on expanding capabilities beyond simple queries to complex task completion and "dreaming phase" assistance, ensuring your agent becomes the customer's preferred interaction channel.

Key insights

Despegar's Sofia agent showcases evolving a multi-agent AI system for complex customer journeys, building custom orchestration before industry standards.

Principles

• Decentralize agent development to domain teams.
• Evolve AI architecture with increasing complexity.
• Combine LLM capabilities with human operational know-how.

Method

Begin with direct tool calls, then introduce an orchestration layer for satellite agents. Externalize agents from the core orchestrator, establishing inter-agent communication protocols (MCP).

In practice

• Integrate AI search toggles into existing platforms.
• Explore in-trip use cases like personalized recommendations.

Topics

• Despegar
• Sofia AI Agent
• Multi-Agent Systems
• AI Orchestration
• Travel Tech
• Customer Experience

Best for: CTO, VP of Engineering/Data, AI Architect, AI Product Manager, AI Engineer, Director of AI/ML

Related on AIssential

• Building an AI-Powered Contract Price Adjustment Agent with Docusign MCP and Microsoft Foundry — The Model Context Protocol (MCP) enables AI agents to seamlessly integrate with and orchestrate enterprise services.
• A New Kind of Marketplace — Building trust and a robust "harness" infrastructure are key to enabling disruptive, agent-driven marketplace experiences.
• 🥇Top AI Papers of the Week — Advancements in AI agent systems focus on structured evolution, efficient reasoning, and dynamic collaboration for improved performance and scalability.
• Building AI Agents: A Journey with Sam Bhagwat — AI agent development requires understanding core concepts, organizational design, and deep integration into customer processes for success.
• Developer’s Guide to AI Agent Protocols — Standardized protocols simplify AI agent development by eliminating custom integration code across diverse systems and user interfaces.
• Tripadvisor, Airbnb: Can AI Agents Beat Travel Agents? — AI agents are transforming travel planning by offering personalized, conversational experiences that leverage proprietary data and LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MLOps.community.