I had a business idea, and a crew of AI agents built the entire platform for me. The website, the iOS app, the outreach, the SEO, the tracking, the analytics. Everything.
This is not a demo of a tool. It is a record of what actually happened when I stopped managing tasks and started running a crew: a harness that multiplies what a single person can do. Watch the video on YouTube. In it I show you everything the crew built, the blind spots it could not see, and what my job looks like now, because it is not what I expected.
The idea
The idea fits on a napkin. Audio tours of Madrid, narrated straight to your phone. A day pass for those who have just arrived and will be gone in a few days. A month pass for those staying a little longer, who want the freedom to choose when to walk and discover the city. A group pass for families and friends who want to do it together. The first stop of every tour is free, a teaser and a trial. That is the whole business.
What the crew turned it into
Seven tours of Madrid. Eight languages. 643 pages. The website you are looking at, with payments and checkout, takes real money from customers. An Android app and an iOS app, a six-screen pricing flow under 3 megabytes, with the tour audio streaming from the server the moment a payment lands.
The SEO: every stop page was structured for search, the sitemaps were submitted, keywords are tracked, and Google search placements are verified on a daily basis. It even checks each day for the errors Google has flagged and fixes them automatically.
The tracking: every visit, every stop played, every checkout attempt is recorded and reported. Without it we would never know whether anything was working, or, more importantly, failing.
The outreach: personalized pitch emails to travel bloggers, discount codes for them, revenue-sharing agreements, all written out and logged by the very same agent loop that wrote the tour scripts. A single tour stop needs research, a script, factchecking, a cross-stop revision pass, translation, voice audio generation, and finally deployment. Seven tours in eight languages is that cycle, hundreds of times over.
And the whole infrastructure bill is negligible. Local models, a few cents of text-to-speech, Cloudflare’s free tier. This used to be a six-month agency engagement with about eight people. This crew did it in about three weeks, and it was just me reviewing.
You would be forgiven for thinking the machine did everything. It did almost everything. Almost is the most important word in the whole video.
The four blind spots
There were blind spots. Let me name four, because every one of them was invisible to the crew, not because the crew failed, but because of the kind of thing it was.
1. The voice
We cycled through three text-to-speech providers before we settled on the right one. The right one speaks English with a Spanish accent, because you are in Madrid, and that is the note I wanted to hit. But the two before it produced audio that was technically perfect and emotionally dead. It is like a museum guide who has read the room and chosen despair. Every spec was met. But “sounds like a person who wants to be here” is not a spec you can specify. I had to make that call by listening to the audio.
2. Repetition
Every tour stop’s script was written independently. So the tour repeated itself. The word ka was explained at five different stops. Churros were pitched seven times, as if there were nothing else in Spain but churros. At one point the narration announced the “final stop” three stops too early. Each unit was correct in isolation; the whole, when you looked at it as a whole, was a mess. No single agent had the entire tour in view. I did. That is the role I had to play.
3. The order
The stops were sequenced by latitude and longitude. Neat, logical, and wrong. The actual walking order depends on the path you used to get there, and it is usually different. Take a food tour: the recommendations should have begun with an appetizer and ended with a post-dinner drink. That wraps things nicely, does it not? But the restaurant with the best appetizer was stop number three, as it came out of the agent. The map answered the questions perfectly. It was simply the wrong question.
4. The one only a body can find
This is my favourite. A user starts the tour wherever they are, and when they get near a stop the audio begins to play. It sounds fine, right? It should be fine. But actually, no. Because someone spends ten minutes walking to a stop with their earphones in, and it has been blank the whole way, and then suddenly, when they approach, the narration blares. It shocks them. The fix is rather simple: play background music while they walk, and let the narration slowly fade in as they near each stop, so there is no shock. No code review, no fact check, no automated test could ever have caught that. You find it only by walking the tour yourself, on a real street, with your headphones on.
Look at the pattern in those four. Taste. Seeing the whole product. Testing in the real world. Factoring in lived experience. Not one of them is a typing problem. The crew did the typing flawlessly, but it still needed my judgment to direct it.
What my day actually looks like
Here is my morning. I open the dashboard. The crew has already reported what happened overnight: what was visited, what was played, where people dropped off, which outreach got replies, what broke. I use my experience and my intuition to suggest a fix, or to try an entirely new idea. And I have the immense pleasure of thinking about ideas without ever worrying about the constraints: the effort it would take, whether I even have the technical capability to build it. That is starkly different from when building used to require eight people and a budget, and you were afraid to step out of the lab for fear all that time and effort would fail to deliver the thing.
The crew treats every idea as a hypothesis. It builds the implementation, puts it in front of reality, and measures what comes back. If something works, it doubles down on it. If it does not, it shelves it fast.
So I stopped managing tasks a long, long time ago. Now I just evaluate outputs. The machine is the multiplication. My judgment becomes the multiplier. And that is the shape of my day: look at reality, form an opinion, then let the crew test it. That is it. That is the job. And it is the best job I have ever had. I cannot believe it did not exist three years ago.
What the product actually is
The product is live. Day passes, month passes, group passes, and the first stop of every tour is free. So if you are in Madrid, go walk with it: reddwalks.com.
But honestly, the product is the proof of concept. What I actually want to show you is the crew itself, the business harness that takes a person’s ideas and runs them as experiments until something succeeds. That is the part that interests me most. It is called Redd Crew, and next week I will take it apart and show you exactly how it works, so that you can build your own. It is live here.
Building is cheap. Judgment is not.
Discussion