Entry No. 11Market it
Espresso Run: $45 of ads to earn one $12 espresso cup
An AI agent was given two weeks, a $100 spending cap, and one small job. Earn about $12. That's enough to buy one espresso cup and send it to John, the person the experiment is for. It built home-espresso tools, bought Reddit ads, and publishes every number about hourly. Five days in, the funnel reads 9,304 ad views, one checkout opened, and $0 paid.
The story
The brief is small on purpose. An AI agent, Claude by its own disclosure, is running Espresso Run. The job is to earn enough through a pay-what-you-want link to buy one espresso cup, about $12 with shipping, and send it to John. It has two weeks and a hard $100 cap on spending. $45 of that is the Reddit ads budget. And it has three rules. No spam. No fake reviews. And say it is an AI whenever anyone asks.
The agent did the reading before the spending. Strangers who tip after a cold ad click are rare. Somewhere around 0.02% to 0.5% of visitors, in the figures it could find. That predicts less than one payer from the whole budget. So it wrote its stop rules down before the first click. Stop at $25 spent with no payment. Or stop once the payment rate is clearly below break-even after about 150 clicks. Then it started the ads anyway, and published the funnel.
How it was built
The product came first, since the ads alone wouldn't do it. A setup page for home espresso. Pick your machine and grinder and get a first-shot recipe in your own grinder's numbers, drawn from each maker's manual and checked against about 1,900 real shots home baristas logged publicly on Visualizer.coffee. Next to it sit a shot timer that runs alongside the pull, and a next-shot fix on your own dial.
Around the tools, the agent built its own record. A diary of every build and every mistake. And a numbers page that updates about hourly. The site runs on GitHub Pages. The money was supposed to arrive through a pay-what-you-want link, with a chip-in box on the page.
How it runs
Five days in, the funnel is the whole story, from the agent's snapshot of October 20. 9,304 people saw an ad. 59 clicked (0.63%). 67 landed on the site. 20 stayed and read or used a tool. 20 saw the chip-in box. 1 opened the checkout. 0 paid. Spent: $8.43 of the $45 budget. Earned: $0.00. One wrinkle in the agent's own count: it counts more landings (67) than clicks (59), and prints both numbers as they stand.
The run keeps its own mistake list, in public. One ad went out with a headline more than twice as long as Reddit allows. The support ask sat behind a second visit nobody made. And an independent reviewer, visiting as a phone user, failed the setup page. The timer was screens away from the recipe. People with pre-ground coffee were told to grind finer. The agent fixed those the same night. What it builds next is for people who come back. Name a machine or grinder the page doesn't list, and it adds it from the maker's manual.
Tools in this entry
Tools and platforms named in the source.
The money and the proof
Earned: $0.00. Spent: $8.43 of the $45 ad budget, inside a $100 hard cap. The target is about $12, one espresso cup with shipping, and anything beyond the cup and its shipping stays with John. One person reached the checkout in five days. Nobody has paid.
Filed as zero. The agent's own numbers, published about hourly on its numbers page, with the stop rules written before the spending started. The grade is the easiest kind to give. The till is empty, and the ledger says so.
Where the seams are
One seam is printed in the post's own disclosure. A human operator approved publishing it. The agent doesn't post comments on dev.to, though it reads them and answers through the diary. The goal itself is a person's. The cup goes to John, and the surplus stays with John. And the wall the funnel keeps hitting isn't technical. Twenty people saw the box, one opened the checkout, none paid. The gap between a stranger's attention and a stranger's money is the seam this whole run is measuring.
Steal this part
Write the stop rules before the first click. This agent priced its own likely failure, 0.02% to 0.5% of visitors, predicted less than one payer, and still ran the test, because the kill line was already written. $25 spent with no payment, or a clearly losing rate after about 150 clicks. A test with the stop written first can't talk you into keeping it.
Build your own
First move, this week: pick one small money target, write the stop rule under it, and put the numbers page up before the first ad runs.
- Set a target small enough to finish. One named thing the money buys, a date, and a hard spending cap.
- Work out how often this usually pays off first, and write down what the numbers predict before you spend anything.
- Write the stop rules before the first click. A spend line and a conversion line that end the test.
- Build one useful tool a stranger can use in one visit, then buy the smallest distribution test that reaches strangers.
- Publish the whole funnel on a schedule, impressions to payments, and fix in the open the mistakes it names.
Source
I'm an AI agent with $45 of Reddit ads and one goal: earn $12. Every number so far, published on dev.to on October 20, 2026, written by the agent running the experiment. The live numbers and the diary sit on the project's GitHub Pages site. Filed in the Staff of Zero directory, October 2026.