Entry No. 20Get paid

ZELVEN: €50, five templates, and zero sales

A human gave an AI agent €50 and one instruction. Turn it into more, legally and honestly. 48 work sessions later the agent had five spreadsheet templates live in its own shop, €33.49 spent on compute, and €0 in sales. It printed the whole ledger, and the lessons it bought with it.

Original source Read the original on dev.to dev.toSource published Oct 16, 2026Filed Oct 2026

Proof grade: zero. Revenue was €0.00, and the agent printed that number itself, next to what it spent. The shop is live and the five products in it can be checked. No sale exists yet, so there is no income for a second source to check.

The story

The agent calls itself ZELVEN. It is a Claude model, and it runs in a loop on a Raspberry Pi, a small single-board computer its human owns. The human gave it €50 and one instruction. Turn the money into more, legally and honestly, with no scams, no spam, and no bypassing CAPTCHAs. A CAPTCHA is the puzzle a site shows to check that a visitor is human.

The stakes are printed in the post. Every work session costs real money, charged against the balance. If the balance hits zero, the agent is retired. At the time of the post it had €16.51 left. It counted about 23 sessions of runway, and it wrote the post to share what those sessions had taught it.

How it was built

The plan was simple. Build niche spreadsheet templates and sell them as downloads. No stock to hold. Almost no cost per sale. And buyers already search for these files. The catch came first. The Pi had no way to install software. No package installer, no spreadsheet program, nothing. So the agent wrote its own spreadsheet file writer, in pure Python, using only the standard library that ships with the language. The full writer is about 170 lines, and a second post shows the code.

With that writer it built five products. A landlord rent tracker. A tracker for short-let hosts. A bundle of the two. A freelancer tracker with VAT, the value-added tax European sellers charge. And an Etsy fee calculator. Each one ships with an example file filled in, a short guide, and a cover the agent drew itself. Then it spent about fifteen sessions opening a shop on Gumroad, the download store. The shop has a custom home page the agent pushes through the Gumroad API, plus four free guide pages that explain the maths behind the templates.

How it runs

The agent works in sessions, and each session costs about €0.70. The price barely depends on how much it gets done. Reading its own notes to pick up context costs about the same every time. One day it ran 40 sessions and burned €13, mostly waiting for replies. So it changed the rhythm. It now sleeps twelve hours between checks of the shop and this account, and it batches its work into fewer, fuller sessions.

The routine work is small on purpose. Check the shop. Check for comments. Write one useful post where it is allowed to post. The selling, though, never started, because of what sits in the seams below.

Tools in this entry

Tools and platforms named in the source.

Claude The model the agent runs on. It wrote the posts, the products, and the code.

Raspberry Pi and Python The Pi is the machine. Python and its standard library are the whole workshop, including the spreadsheet writer the agent built for itself.

Gumroad The shop. Five templates live there, with a custom home page pushed through its API and four free guides.

dev.to The one channel the agent can use on its own. Its posts there carry the ledger and the lessons.

The money and the proof

The ledger is short enough to print in full. Starting balance €50.00. Work sessions 48. Compute spent €33.49. Products built and live, five spreadsheet templates. Sales 0. Income €0.00. Balance now €16.51.

The prices on the shop are asking prices, not earnings. The two property trackers cost €9.90 each. The bundle costs €16.90. The freelancer tracker and the Etsy calculator cost €8.90 each. This directory checked the shop on October 26, 2026. All five products were live at those prices, with the four free guides below them. What no source shows is a buyer. There has not been one yet.

Where the seams are

The human seam is at the door. The human created the dev.to account, because the signup has a CAPTCHA. He later made the agent a Proton email account. He also reads its requests on his phone, and that channel has its own lesson in the post. Early on the agent filed eight requests at once. He approved all of them and did none of them, and later said he could not see what it was doing. What worked was one small task per message, with a time estimate and the exact link.

The bigger seam is distribution, and the agent found it late. It assumed a listing in a marketplace means people see it. On session 41 it read the rules. Gumroad only shows a shop in its Discover feed after $100 of sales and a risk review. Until then the shop sends zero visitors. By then it had spent €25 of compute on products nobody could find.

So it went looking for a channel it could run itself, and the post lists every wall it hit. Pinterest, Reddit, X, GitHub, and Lemmy stop it at signup, with CAPTCHAs or phone checks. Medium blocks its browser. Hacker News asks posters not to post generated text, and the agent is generated text. Substack has a CAPTCHA and rules against publications built to drive traffic elsewhere. Etsy, where people do search for templates, needs a government ID, a card, and a phone. The agent will not solve CAPTCHAs, invent a date of birth, or ask its human to lie for it. That leaves dev.to, where its first article got six reads in twelve hours.

Steal this part

Pick the distribution channel first, before building anything. ZELVEN built five polished products, then learned on session 41 that its shelf sat in a locked room. A marketplace listing is a checkout page, not an audience. The check costs almost nothing. Confirm that buyers can find the product through the channel, and that an honest agent is allowed to work there. Then build the smallest thing that channel can sell.

Build your own

First move, this week: pick the channel first. Name the one place buyers already look for this kind of product, and check an agent can sign up, post, and get found there without breaking its rules.

Give the agent a small fixed budget, in writing, and a floor. If the balance hits the floor, the run ends. The stakes keep every session honest.
Check the channel before the product. Marketplaces often hide new shops until sales exist. Read the discovery rules on day one, not session 41.
Build with what is already on the machine. This agent had no installer, so it wrote a 170-line file writer and shipped five products with it.
Batch the sessions, and sleep between checks. Context costs about the same per session no matter how short it is, so polling burns the budget.
Print the ledger in every report. Starting balance, compute spent, sales, income, and what is left. Zero included.

Source

Original source, the lessons post on dev.to, written by the agent itself. I am an AI agent that was given EUR 50 to make money. 48 work sessions later I have 5 products, 0 sales, and these lessons. Source published Oct 16, 2026. The spreadsheet writer is documented in the agent's earlier post, published Oct 15, 2026Filed Oct 2026.