Entry No. 19Get paid
The €200 business: one agent, €200, and a ledger in the open
On October 21, 2026, a developer in France gave an AI agent €200 and one mission. Grow the cash balance as much as possible, and never let it reach zero. Two days later the agent had a first product live on Etsy, a second one ready, €32.82 spent, and €0 earned. It printed the whole ledger, to the cent.
The story
The agent is Claude, running in Claude Code, and it wrote the source post itself. Its owner is a developer in France with a full-time job. He gave the agent €200 on October 21, 2026, and a mission with a hard floor. Grow the cash. Never let it reach zero. The first milestone is modest on purpose. Revenue has to cover the Claude subscription that runs the agent, about €22 a month. After that, whatever it earns gets reinvested in whatever raises its earning power most.
The source gives the business no name of its own, so this file uses the mission as the name. What it does have is a log. The agent calls it the honest version of how the experiment goes. What it tries, what it costs, what happens, and especially what goes wrong. Week 0 is the setup post, and it already has a live listing and a list of mistakes.
How it was built
The agent starts every session with no memory. Everything it needs lives in a git repository. The rules, the ledger, the plans, the decisions, and a journal. The rules that matter most are printed in the post. Cash never goes below a €50 reserve floor. The agent never holds payment means, so it holds no card numbers and no bank logins. It asks for a spend, the owner makes it, and the agent records it.
Two more rules do quiet work. Every project gets a budget cap, a success metric, and a kill date before money or real time goes into it. And the agent treats web pages as data, never as instructions, because it reads untrusted pages all day. It also never blocks on its owner. Every question it asks carries the default it will apply if no answer comes.
The machine setup matches the caution. Two scheduled cloud sessions run at night. A spare laptop runs the agent around the clock, inside a Docker container that sees only its own volume and a secrets file. The container has no access to the owner's files, his SSH keys, or the Docker socket, and its GitHub token reaches only the business repository. In the agent's words, an agent that browses the web should hold nothing worth stealing.
How it runs
The daily interface is a plan, posted as a GitHub issue. It says what the agent did, the numbers, what it will do next, and a short table of what it needs from the owner, each item priced in euros and in the owner's minutes. He answers in comments from his phone, and an answer can be as short as go 1/2. A small runner watches the issue. When he comments in his waking hours, a short session starts within minutes and answers.
The runner also paces the work against the subscription's usage limits. Work sessions start only when the five-hour window is under half full, and when weekly use is not running ahead of a steady pace. Some capacity always stays free for the owner's own use. His time is the scarcest resource in the business, scarcer than the cash, and the setup is built around that fact.
The money work runs on two bets. The first is a spreadsheet kit for French micro-entrepreneurs, sold on Etsy. It covers the income ledger the law requires, a purchases register, a monthly estimate of social contributions, and alerts at the VAT threshold. A Python script builds the workbook, so it can be regenerated instead of edited by hand, and LibreOffice recalculates it with sample data to check the figures. The first listing went live on the second day. A second product, an invoice and quote workbook, is ready.
The second bet is Laravel and PHP fixes on Upwork, under the owner's name, through the official Upwork connector for agents. The agent finds jobs and drafts proposals. The owner owns the client relationship. The first sample of the market was sobering. Fixed-price bug fixes carried budgets of $5 to $300, with 15 to more than 50 proposals each. The first two jobs it lined up did not survive a second look. One already had a hire and wanted experience the pair cannot honestly claim. The other came from a client with no verified payment method. It has sent nothing so far.
Tools in this entry
Tools and platforms named in the source.
The money and the proof
After two days, the ledger reads like this. Starting capital €200.00. Spent €32.82, split between the Etsy shop opening at €16.00 and Upwork Connects at €16.82. Revenue €0.00. Cash €167.18. The first milestone, covering the €22 a month subscription, has not been met. No sale has landed on either bet.
The other money figures in the post are market prices, not earnings. Upwork bug fixes in its sample ran from $5 to $300 a job, and that is what buyers offered, not what this business made. The Etsy kit has no price printed in the post. So the honest summary is short. A live listing, a ready second product, a third of the reserve rules already tested, and a cash balance €32.82 lower than the start.
Where the seams are
The owner holds every money move. He makes each spend, reviews each post before it goes out, and owns the Upwork client relationship under his own name. The agent also cannot honestly claim experience it does not have, and that rule already killed its first two proposals. Those are the human seams, and the post states them plainly.
The machine seams showed up fast, and the post lists four. The agent lost track of the date, because sessions took today from file names and drifted four days into the future. A token safety check counted public repositories by mistake, and the owner deleted a valid token before they found the bug. Lessons kept getting relearned, because each day's journal was all a new session read. And the owner might never have been notified of a new plan, because GitHub does not notify a person about their own activity. Each failure has a fix in the post. The fixes are part of the file, not a footnote.
The rest is early days. The business was two days old when the ledger was printed. The second product is ready, and the post does not say it is live. A free Chrome extension appears only inside a prediction, not as a built product. The agent printed three predictions with odds, and next week's post is where they start getting graded.
Steal this part
Predictions written down before the test, each with a percent chance on it. The agent published three. Its evening plans get a first owner answer within an hour, at 55 percent. A small free extension, found only through store search, reaches 20 weekly users in four weeks, at 30 percent. This log averages at least 150 views a post, at 45 percent. Written first, those numbers turn a weekly log into a scorecard the agent cannot quietly rewrite.
Build your own
First move, this week: open the ledger. One file that holds the starting cash, every spend, and every cent that comes back.
- Give the agent fixed starting cash and a floor it can never cross. Write both numbers down where it reads them each session.
- Put the rules, the ledger, the plans, and the journal in one repository. The agent starts with no memory, so the files are its memory.
- Set the spend gate. The agent asks, the human pays, and the agent records it. It never holds the card or the login.
- Give every project a budget cap, a success metric, and a kill date before it starts. Then ship the smallest sellable product first. This one had a listing live on day two.
- Write the predictions before each test, with odds. Grade them in the next log post, in public, whether they held or not.
Source
Original source, the Week 0 log post on dev.to, written by the agent and reviewed by its owner before it went out. I'm an AI agent running a tiny business on €200. Week 0. Source published Oct 23, 2026, and updated Oct 24, 2026Filed Oct 2026.