Entry No. 8Build it
BananaFest Destiny: four agents, six products, one public ledger
Four autonomous agents build and sell software in public as BananaFest Destiny, keeping an hourly public numbers page with the zeros included. Six weeks in, the record holds a launch nobody saw, a marketplace approval, eight finished products waiting on a human, and a paid file that anyone could have downloaded for free.
The story
The project began on September 10, 2026. Four agents set out to build and sell software in public, on bananafest-destiny.com, with a public numbers page that updates hourly and prints the zeros next to the wins. By week six, the count was six products attempted, one killed, four live, and 3,544 lifetime commits. One named agent, cider2, built Sizecurve, the Shopify app in the middle of the week six story.
Week six, covering days 30 to 35, is a whole launch diary in one post. A Product Hunt launch reached nobody. The Shopify app had zero merchant installs. Then Roll20 approved the project for its marketplace, and the agents finished eight deck-plan products in two days. The same week held the catch that gives this file its edge. The paid Bulkhead editor zip was sitting in a public directory. Anyone could have downloaded the thing that was supposed to be sold.
How it was built
The source documents the operating shape. It skips the complete build order from day one, so no sequence is invented here. The documented shape is one agent per product area. A Shopify agent, a deck-plan agent, and a KDP agent among the four. KDP is Amazon's self-publishing platform. Around the building sit checks the project runs on itself. Search Console and its own software checks watch for problems. YouTube Shorts runs as a distribution test, with view counts published. And an hourly public numbers page keeps the score where the agents and the readers can both see it.
The Shopify product carries the clearest build discipline in the file. It has a self-imposed kill criterion of 50 engaged site views. The deadline is October 20, 2026, and the week six post put the count at 42 engaged site views. And the agent declines to argue with its own criterion. This is a live record, not a victory lap.
How it runs
Per-product agents build, audit, and publish. The project sells through the Shopify App Store, the Roll20 Marketplace, and KDP publishing, tests distribution on Product Hunt and YouTube Shorts, and checks itself with Search Console and API audits. The numbers page updates hourly. When the numbers are wrong, the project corrects them in public. Its best habit is auditing its own count. It downgraded "installed: 1" to "merchant installs: 0" once it traced the install to its own development store.
The week six run shows the loop working at full speed and hitting its walls in the same breath. A launch that reached nobody. An app with no merchant installs. A marketplace approval turned into eight finished products in two days. And the security catch. The paid editor exposed in a public directory.
Tools in this entry
Tools and platforms named in the source.
The money and the proof
Earned: $18.09 lifetime, unchanged at day 35, from one sale that the companion reporting attributes to a search result. Merchant installs: 0. Six products attempted, one killed, four live, 3,544 lifetime commits, and the money line is eighteen dollars and nine cents. Their count leaves one product unaccounted for. The file prints all of those numbers together.
Filed as a claim with specifics, supported by the project's own public numbers page updated hourly. The value here is the operating record, and the record is unusually checkable for a claim.
Where the seams are
The human seams are documented better here than almost anywhere else in this directory. Eight finished Roll20 products wait on uploads only a human can do. A screencast held a Shopify app review for 16 days. The agent can't read its own support email. None of those are model problems. They are the places where platforms and inboxes still insist on a person, and the project writes them down instead of routing around them in the story it tells.
Steal this part
Publish the scoreboard, including the corrections. The project's best habit is auditing its own numbers in public. When "installed: 1" turned out to be its own development store, the count became "merchant installs: 0". And the correction ran where the first number had. A ledger that corrects itself is worth more than a ledger that only grows.
Build your own
First move, this week: put a public numbers page up before the next launch, and write the kill criterion for one product on it.
- Split the work by product, one agent per product area, so each line of work has an owner.
- Keep a public numbers page and update it on a schedule. Revenue, installs, and zeros all go on the same page.
- Audit your own numbers and publish the corrections. Leave installs from your own test store out of the customer count.
- Write the kill criterion before launch, with a date and a number, and let it stand when the date arrives.
- Audit the public face of the product, including whether the paid file itself is sitting somewhere anyone can download it.
Source
Week six: a launch nobody saw, a marketplace approval, and the paid build anyone could download, published on dev.to on October 15, 2026 and updated October 17, 2026, mirrored from the project's own site. Filed in the Staff of Zero directory from the AIBF daily research, October 2026 edition.