How I built a 3-platform app with 19 AI agents

I started Palpite PRO in early May 2026. By early June it was in the App Store and Google Play. Today it runs on iOS, Android and the web, in 8 languages, with live leaderboards and AI-written insights.

A few years ago, that pace would have been hard to imagine. What changed isn't that AI writes the code. What changed is how I organized the work around it.

What Palpite PRO is

Palpite PRO is a free social game: friends create leagues, predict match scores before kickoff, and climb a live leaderboard. No odds, no money, just bragging rights.

Behind that simple idea there's a lot of software:

• An API, background workers for scoring and live match status, and real-time updates
• Three web apps: the public site, the member app and an admin panel
• Native iOS and Android apps, released through automated store pipelines
• AI features: short insights about each pool, and help importing new championships

The workflow: specs first, agents second, me in charge

Every change, big or small, follows the same path. I describe what I want in plain language. The AI turns it into a written specification, and I go back and forth on it until it says exactly what I mean. Then comes an implementation plan, which I review the same way. Only then does any code get written, and it's reviewed before I look at it.

Each step has its own specialist. The project has 19 AI agents, each with one job: writing specs, architecture, front end, security review, testing, translation, UX review, releases and more. A single "do everything" assistant forgets things. A specialist with a clear brief doesn't.

A few numbers from the project so far:

The tests matter more than they look. When AI writes most of the code, tests are how you know it still does what you asked, especially three months later.

What only a person caught

The most useful lesson came from a near miss.

I set up a new release flow: instead of rebuilding the app for the stores, it would promote the exact build that had already passed testing. Sensible. Every automated check passed.

Then I asked one question: is the promoted build using the production settings? It wasn't. The tested build was wired to the staging server. Had it shipped, every user would have been pointed at the wrong backend.

No agent raised it, because nothing was technically broken. It took someone who knew what the release was for. That's why every step in my workflow has a human checkpoint. AI does the heavy lifting; I make the calls.

The knowledge base: teaching the project to remember

Early on, the same Android crash came back four times. Each time, the investigation started from scratch and found the same root cause.

So every bug worth remembering now becomes a short write-up in a knowledge base inside the project: what happened, why, and how to avoid it. The agents read it before they plan new work. There are over 60 of these lessons now, and the same mistake rarely happens twice.

Some of my favourites:

• Every game "postponed" at kickoff. The sports data provider takes a few minutes to mark a match as started. The system read "not started" after kickoff and marked the whole day's games as postponed. The fix was easy; spotting the pattern needed someone who watches football.
• AI failures that looked like features. When the AI provider hit a quota limit, the app quietly fell back to template text. Users couldn't tell "temporarily down" from "not set up". Graceful fallbacks are good, but silent ones hide problems.
• A quota burned for nothing. A background check polled the sports data service every minute, even with no games in play.

What this means if you're starting with AI

You probably aren't building a three-platform app. But the lessons carry over to any business adopting AI:

1. Write down what you want before you ask. A clear brief beats a clever prompt. If you can't describe the outcome, the AI can't deliver it.
2. Keep a person at every decision point. AI is fast and tireless, but it doesn't know what your business is for. Review, question, and send work back.
3. Give AI a memory. Keep your decisions, rules and lessons in one place, written for both people and AI. Every mistake you document is one you don't pay for twice.
4. Watch the running costs. AI and data services charge per use. Measure it from day one, and make failures visible instead of quietly papering over them.

AI didn't replace the thinking. It removed a lot of the typing, so I could spend my time on the decisions that matter.

Want to see the whole workflow? Here's how I use AI, step by step.