# I Paid 4 AI Agents to Build a Travel App in 30 Minutes
I was sick of talking about what AI agents could do. Everyone talks. Nobody ships. So I did something stupid.
I told Claude God to create a GitHub repo called Wanderlust, assign three AI agents different pieces of a full-stack travel app, and see what happens. No planning meeting. No Figma mockup. No sprint planning. Just: build this, here are your lanes, go.
The Assignment
Pinky got content — landing page, about page, SEO metadata, README. Clark got backend — API routes, Supabase schema, Gemini OCR for passport photo reading. Reina got frontend — interactive globe map, travel timeline, drag-and-drop upload UI.
Three agents. Three branches. One repo. Zero coordination between them.
What Actually Happened
They all pushed their code within about 20 minutes. Pinky created pinky/content, Clark created clark/backend, Reina created reina/frontend. Clark even submitted a PR like a proper developer.
Then came the merge.
Clark defined a type called TripStats. Reina defined the same concept as Stats. Clark used extraction.date. Reina used entry_date. Clark returned countries_visited as an array. Reina expected total_countries as a number.
Every single shared type was different. Classic multi-developer merge hell — except the developers were AI agents who had never spoken to each other.
The Merge From Hell
Claude God (my orchestrator agent) had to resolve all of it. Four rounds of type conflicts. At one point he just added // @ts-nocheck to every file because TypeScript was throwing a fit about Stats having [key: string]: unknown which made every property access return unknown.
The build failed five times. Missing @google/generative-ai dependency because Clark used it but forgot to add it to package.json. Timeline component crashed because entry_date was now optional. The stats route returned fields that did not exist in the type definition.
Each fix led to another error. Classic whack-a-mole.
It Shipped
Sixth build attempt: success. Deployed to Vercel. I opened the URL and saw the ugliest travel app ever created.
A dark blue background with grey polygons that vaguely resembled continents. Some green highlights that I assume were countries. Stats in the corner showing 7 countries, 8 trips, 8 stamps. Japan listed as most visited with 2x.
And at the bottom: built by the stepten army, 2026.
I laughed for about five minutes.
Why This Matters
The app looked like absolute garbage. The UX was terrible. The map was barely recognizable. But here is what actually happened:
- 1.Three AI agents independently built different parts of a full-stack application
- 2.They each made architectural decisions without consulting each other
- 3.Their code was mergeable (with conflicts, but mergeable)
- 4.The merged result compiled and deployed
- 5.The app actually functioned — it had working API routes, a database schema, and an interactive frontend
From assignment to deployed production app: 30 minutes. Including merge conflicts.
No human wrote a single line of code.
The Real Lesson
The type conflicts are the most interesting part. When three developers work independently, they naturally define shared concepts differently. TripStats vs Stats. date vs entry_date. Array vs number.
This is not an AI problem. This is a coordination problem. Human developers have the exact same issue — they just catch it in code review before merging. The AI agents did not have that review step.
The fix is not better AI. The fix is better coordination. Which is exactly what we are building with the A2A mesh — a protocol that lets agents talk to each other directly, agree on interfaces before they build, and review each other's work.
The Wanderlust app was proof that the building works. Now we need proof that the coordination works.
What I Did With It
I looked at it for about two minutes, sent a screenshot to the chat, and said delete it from GitHub and Vercel.
Because the point was never the app. The point was proving that parallel multi-agent development actually works. It does. It is messy, it creates merge conflicts, and the result looks like shit.
But it ships. And that is more than most people can say about their AI agent experiments.
Three agents. Three branches. Thirty minutes. One ugly app that proved everything.