Ilbets

Posted on May 24

A New Era of Agent Native Applications

#llm #ai #startup #travel

In the next five years, 99% of companies will have agents for their services. That will open up a new world of agent-to-agent communication—an agent economy and even an internet of agents. Agents may not completely replace traditional apps, but they’ll take a huge market share.

Next time you want a new dining table, it might be as simple as asking your Agent app. It’ll message a few store agents, find one you’ll like based on what it knows about you, and place the order.

So, when I started working on Terminal1.app, I wanted to imagine that future. I asked myself what an “AI native” company might look like—and built it, agent-first, from the ground up.

Idea

In late 2024, I had a three-hour layover in Munich. I was alone, eating dinner, and thinking about how much time I’d wasted in airports that year. I’d taken around 60 flights for work and leisure. That felt like a lot, but some of my friends, especially those in sales, fly much more.

I sometimes try to work, code, or read. But let’s be honest — getting anything done in an airport isn’t easy. I started wondering: how many people like me are sitting in airports right now, looking for something meaningful to do?

That’s when the idea struck me — what if there was something in between Tinder and YC’s Co-Founder Matching to build business connections during travel? An app that connects airport business professionals based on their shared interests in the airport.

Lastly, meaningful social connections are key to personal happiness. I’ve met many people who later became close friends at events and gatherings. Sometimes we’re so busy with work and chores that we forget how important it is to be around people we care about — and who care about us.

Terminal 1

Terminal1 connects traveling professionals at airports based on their LinkedIn profiles, flight times, and interests — so they can network while they wait. Transform airport time into business opportunities.

1️⃣ Log in with LinkedIn to create a profile

2️⃣ Send your CV to create a detailed profile for matching

3️⃣ Share your upcoming flight details

4️⃣ Review and accept your match 24 hours before the flight

Communication in the Agentic Era

One thing that changes in the new Agenting Era is how we communicate with customers. You don’t need fancy designs to turn simple user needs into system actions. Let the agent handle it for you ✨.

The primary communication interface for Terminal1 is email (maya@terminal1.app): send your CV, forward flight tickets, or reach customer support.

While there’s also a web interface, it’s intentionally simple, as it’s not necessary in AI-Native app.

However, I found it useful to address trust issues when working with AI agents.

From the 50s to AI Native

From day one, I wanted to build an AI Native system. But first, I had to decide what that really meant for myself. Since agents act like humans, I imagined what Terminal1 would look like as a company built in the 1950s, where everything is handled by people.

It starts at the reception desk — a scene we all know from old movies. A mailbox, a stack of documents, and a receptionist asking, “How can I help?” They take your request and promise you’ll get a reply by mail in a few days.

They sort each request into the right box. Later, an assistant with a trolley cruises through the office, delivering each box to the right department. Your CV goes to the CV review team. Your flight tickets go to the flight review team. The matching team checks upcoming flights, compares profiles, and finds a good match. Then they pass it back to the assistant, who brings it to customer support. Finally, customer support sends you a letter by mail about your upcoming meeting on your next flight.

And that is exactly what I did.

Design of the Agentic System

Terminal1 is built around a network of asynchronous agents. Everything starts and ends with an agent. The system is designed to be modular, where each agent focuses on a specific function, passing tasks along a dynamic pipeline.

Reception agent — identifies the user’s intent from incoming messages and routes the request to the right agent. It’s the starting point of every interaction.
Ticket agent — extracts, processes, and organizes flight ticket data such as destinations, dates, and airlines, so it can be used in downstream matching.
Curriculum Vitae (CV) agent — parses resumes, normalizes the data, and prepares profiles for efficient matching with other users.
Customer Support (CS) agent — handles incoming support requests, retrieves relevant information, and triggers actions to help the ve user.
Public Communication agent — generates the final response to the user, translating system output into clear, friendly language.
Date Interest agent — understanding what the user is looking for in their match: partnership, investment, or job opportunities, etc.
Date Feedback agent — process feedback from the date
Matching Agent — create the best possible match for the date

There are also a few other agents that support these main flows, executing essential tasks like data validation, formatting, and cross-checking.

Agents

Each agent in the Terminal1 system is built using LangGraph. They interact with the system tools through the MCP, which provides access to databases, matching logic, and communication layers.

Some agents, like the Customer Support agent, are powered by the ReAct framework to support multi-step reasoning and tool calling to resolve more ambiguous or complex requests.

Prompts

Getting high-quality results from AI agents requires a lot of prompt engineering. Along the way, I learned a few key principles:

Order matters — Structure your prompt carefully. Put the most important context upfront.
Examples — Good prompts need great examples. They guide the model’s behavior better than instructions alone.
Chain of Drafts (CoD) — An effective technique to improve quality while keeping token usage low. Iterate in stages.
Models — The same prompt can produce wildly different results across models. Choose your model wisely.

Each agent uses a consistent prompt structure:

<Task> <Examples> <Customer-Context> <Real-Time-Context> <Agent-Identity> <Company-Identity>

User Matching Agent

Perhaps the most intriguing part of the system is how we match users. It all starts when someone sends us their CV. The CV agent parses it, extracts key information, and transforms it into a clean, normalized format.

AI-normalized profile summaries are stored in a vector store. This lets us run similarity searches based on meaningful parameters, not just keywords. The Date Intent Agent does something similar — building an ideal normalized match profile for each user based on their input and context.

The rest of the magic happens in a cron job that runs every few hours. It gets all the unmatched passengers flying in the next 12 hours and groups them by airport. For each airport and overlapping time slot, we will take each potential profile and compute potential candidates based on the “perfect profile”.

This matrix, together with users’ profiles and their date intent, will be passed to the Matching Agent to generate final pairs.

Finally, everything comes together in a cron job. The heavy lifting — grouping users by location and departure time — is handled by an SQL function for efficiency. The result is then fed to the agent, which processes each user group individually to evaluate and confirm the best possible matches.

`Multi-dimensional matching based on time, location, and customer context is hard. Pair matching is even harder. You need to look at what both people care about and what they offer. For example:

VC looking for an AI startup to invest in and an AI Startup Founder.
Recruiter from an AI startup and engineering looking for a new opportunity
Talented singer and music producer `

Each match has to be a win-win.

With more user feedback, we can build an “Elo score” to help prioritize candidates on very popular destinations.

Tech Infrastructure

The whole solution is hosted on AWS.

Tech Cloud

Originally, I vibe-coded the website with the bolt.new, but I quickly hit its limits. So I downloaded the code and switched to just coding in Cursor.