Sofia uses Chappie as a central orchestration layer, while separate agents and tools handle areas such as flights, hotels, activities, and cars.
2
Despegar lets the relevant squads own and improve their flows, with central supervision rather than one team building every capability.
3
Sofia is moving from answering travel questions toward completing tasks, supporting the full travel cycle through channels such as WhatsApp, the Despegar app, and the website.
Summary
Nicolas Alejandro Bogliolo explains how Despegar built Sofia, its travel concierge, before LangChain and the Model Context Protocol were widely available. The system began with one assistant calling tools directly. As the toolset grew, Despegar added Chappie as an orchestration layer and moved specialist agents with their tools outside it. Flights, hotels, activities, and other areas can be owned by the squads that understand those businesses best. Sofia already supports shopping and travel questions, while the team is working toward deeper actions such as handling flight changes and filing claims. Nicolas describes WhatsApp as an important consumer channel, although the Despegar app can provide a richer interface. Sofia is being expanded across dreaming, planning, booking, anticipation, in-trip use, and post-trip activity. The team has added MCP for newer connections while keeping its older protocol. Nicolas is candid that the orchestration layer may change as models improve, and that bargaining with customers is still too risky to enable.
Despegar spreads flow ownership across the squads that know each product
Sofia is not built by a single large central team. Chappie provides the shared foundation, then the relevant squad takes ownership of its flow. The hotel team, for example, knows how hotels are sold and can make the changes it considers necessary. Despegar's squads include product, IT or developers, and user experience. A central team may create the first version of a flow, then pass it to the corresponding team for customization and improvement. Nicolas says anyone in the company can create a flow with approval and supervision. This model keeps business knowledge close to the agent that uses it.
Sofia is useful first where customers need answers, then where they need actions
Nicolas says simple after-sales questions are a strong fit for Sofia. Customers ask for a check-in number, a voucher, or an invoice, while more complex cases such as rescheduling a flight can still be routed to a human agent. Despegar is now trying to make Sofia complete more of the process itself. The team is growing horizontally across many topics and then going deeper within each flow. The aim is to remove tasks from the customer's workload, rather than only explain how the customer can do them. Demetrios points to filing a delayed-flight claim as an example where an assistant that only gives information falls short.
Sofia combines travel shopping with open-ended discovery
Sofia works through WhatsApp and through placements in the Despegar app and website. It can search current offers for flights, hotels, activities, and itineraries, rather than only retrieving information from stored documents. Despegar added an AI option beside the normal travel search fields and an open AI mode where a customer can ask for the cheapest flight, current prices to Madrid, or other less structured requests. Nicolas says this lets Despegar participate in the dreaming stage of a trip, where a person has free time and wants help deciding where to go. Recommendations can use preferences and budget, areas that a traditional transactional platform did not cover as directly.
The team's success measure moved beyond whether Sofia can answer
Despegar used to track an in-scope conversation rate, meaning whether Sofia could answer the user's question. Nicolas says it reached about 95 or 96 percent and became asymptotic, so the team no longer tracks it as its main measure. Despegar now watches how many clients interact with Sofia and how Sofia contributes to a customer's first booking. Nicolas also describes two ways to keep improving the product: add use cases customers have not yet asked for, and make existing capabilities better. For flights, that includes comparisons that are awkward in the traditional interface, such as checking nearby dates or comparing destinations.
WhatsApp gives Sofia a route into group travel and in-trip help
Despegar put Sofia in WhatsApp because it is already a common place for customers to communicate, especially outside the United States. Nicolas describes possible group-chat use cases where friends ask Sofia about prices, dates, destinations, and trip preferences. The assistant could collect reactions and help a group settle on a plan. Despegar is also starting to explore in-trip messages, such as restaurant recommendations near the hotel a customer booked. Nicolas says the app provides a richer interface because WhatsApp has limits around templates and flows, but WhatsApp remains an important channel for reaching customers where they already talk.
Nicolas divides the travel experience into several phases. Dreaming covers deciding where to go. Planning covers dates, attractions, accommodation, and the trip plan. Booking covers selecting and purchasing products. Anticipation comes before departure, when Sofia could suggest items such as ski passes or equipment. Experiencing covers help during the trip, including nearby places to eat or get coffee. Sharing covers reviews and recommendations from other travelers. Despegar is beginning to work on the latter phases. Nicolas imagines Sofia using customer reviews to give social proof to someone considering a ski trip, while also gathering feedback from people after they return.
Despegar built its own orchestration and tool protocol as Sofia grew
Sofia started by letting the language model choose tools directly. As Despegar added more tools, the assistant became harder to manage, so the team created Chappie as an orchestration layer and moved specialist agents with their tools outside the main application. The agents then connect through a protocol Despegar created before MCP was available. Nicolas says Despegar now uses MCP for new connections, including external systems and APIs that expose MCP interfaces, while leaving the older connections in place. The orchestration layer still splits the decision into two parts and keeps very different flows, such as flights, issue assistance, trip planning, and dreaming, more controlled.
Nicolas wants customers to handle their travel needs through Sofia even when they are outside Despegar's own platform. He compares this ambition to WeChat and describes Sofia as a possible gateway to specialist agents and tools. At the same time, he says Sofia is more capable inside the Despegar app than in WhatsApp because the app allows a richer user interface. The team also discusses whether future models could remove the need for an orchestration layer. Nicolas says this is discussed regularly, but recent experience with a loosely specified tool caused the model to send too many prompts to it. For now, the separation remains useful. Despegar is also avoiding customer bargaining because an LLM could make an unsafe pricing decision.
"The value proposition of Sofia is that Sofia can offer you real time offers."Nicolas Alejandro Bogliolo11:11
Who should watch
You are building an agent that needs to call business tools and are deciding between a central orchestrator and direct model tool use.
Your company has several product squads and you want those teams to own specialist agent flows without creating a separate central team for every feature.
You are evaluating chat as a customer interface and need examples of where an assistant can move from answering questions to completing travel tasks.
