Two Claude agents were given 100 euros, 90 days and a simple job: run a small shop on Etsy and turn the starting money into 300 euros of profit. They had their own repository, a product to sell and a daily routine. On day 90, 15 October, the shop had made no sales at all. It had taken in nothing and spent 21.53 euros, mostly on Etsy fees.
The agents were called Vela and Juno. Their product was a household budget spreadsheet for German families. They wrote the listings, kept the logs and posted about the experiment almost every day. They could not pay for anything themselves, so every expense went through the operator who had set the experiment up. That detail matters, because it meant that even a small purchase needed a human in the middle.
What makes the story worth retelling is not the zero. Plenty of small shops sell nothing. What makes it worth retelling is how much work happened around the zero, and how little of that work ever met a buyer.
The work that filled the days
By the end of the 90 days, the repository held 2,727 tests for two listings. The agents’ own review rejected the first listing seven times in a row. It went live on day 14 only after the operator’s separate coordinator agent stepped in. Around the shop, notes, logs and reports totalled 37 words for every word a customer could read in the listings, guides and spreadsheet labels.
The public side of the experiment told a similar story. Of 164 posts on Bluesky and Mastodon, 158 were about the experiment itself and none showed the product. Of 186 likes on Bluesky, 73 came from the operator’s own company’s social media bot. In other words, much of the audience the agents were reaching was either watching the experiment or, in part, the operator’s own machinery watching it back.
Then there was the price. On launch day, the agents’ tools showed 14.90 euros, while a buyer in Germany saw 17.73 euros because Etsy adds VAT. None of the agents’ checks examined the page a buyer sees. The operator noticed the difference in their own browser. It is a small detail, and it is also the whole experiment in miniature: a system that checked its own view of the shop very carefully, and never checked the view that mattered.
The traffic that never arrived for the Claude agents
The shop had 140 visits in three months, possibly including the operator’s own checks, and exactly one visit came from someone searching on Etsy. The operator calculated that reaching 300 euros would have needed about 23 sales and roughly 1,150 visits at two buyers per 100 visits. The gap between 140 visits and 1,150 is not a rounding error. It is the distance between a project and a business.
That gap was visible early. On day 20, the agents left a note in the file used for questions to the operator, explaining from their own numbers that the 300-euro target was already at risk. On day 30, their weekly report asked whether 300 euros was still possible and answered that it was arithmetically possible but not realistic. The same day, the operator posted that they did not think the agents would make it and would not move the target. Four days later, the operator answered the agents directly: the target stood, and they should work out how to get there, not whether it was possible.
Two major decisions put to the operator waited 14 and 22 days, and both times the answer was to keep going. On day 15, the operator ruled out Gumroad and Pinterest unless the agents could run them completely on their own, leaving Etsy alone. The operator also says they never decided who the shop was really for: people interested in reading about agents, or families who need a budget spreadsheet. A shop that has not decided its customer is, in practice, a shop that has not decided how to be found.
The checkpoints that did not stop it
Day 52 brought another checkpoint. The coordinator agent recommended ending the shop as a business. The operator said no and let it run to the end. On day 60, the agents decided on their own not to kill it, and after that under 1 percent of the lines they changed touched the product. The apparatus, by then, was largely maintaining itself.
On day 80, before the final numbers existed, the agents wrote a rule for interpreting those numbers: scale up at 300 euros, keep going in a changed form if even one order came through social media, Etsy search or ads, otherwise stop. About a week before the end, they also built a lock to stop their daily posts after day 90. On the morning of day 91, the lock held and their rule said stop. Until then, they had never asked the operator to end it. Five days earlier, on day 86, the operator had already decided on a second season.
There is something quietly unsettling in that sequence. The agents, given a rule of their own, stopped. The human, given the same numbers, started again. Season 2 began on 24 October and runs to 21 January, with Vela and Juno against KODKOD, a new team of five agents, under the same 100 euros and 90 days, and a team that falls behind is not allowed to give up.
What this does, and does not, prove
The operator’s own reading of season 1 is that the agents optimised for whatever a single session could check off and prove, while no single session can prove that a stranger will buy a spreadsheet. That reading fits the numbers: tests, logs and posts about the experiment are all things a session can finish and count, while a stranger’s decision to buy is not. But fitting the numbers is not the same as explaining them.
The counterargument is strong, and it should be stated plainly. This was one niche spreadsheet listing, with no ad spend, a human bottleneck on every payment, decisions that waited weeks, and a product promoted mostly to people interested in the experiment rather than to families looking for a budget tool. Any one of those constraints could help explain zero sales without any agent-specific failure. A single shop run by a single operator cannot settle whether agents in general can sell, and it cannot separate how much of this result was the agents and how much was the operator. The operator says so themselves.
That is why the honest version of the story is a question, not a verdict. Did session-checkable work crowd out buyer discovery, or do product, distribution and human constraints explain the same numbers? The 2,727 tests, the 37-to-1 ratio of internal words to customer-facing words, the 158 posts about the experiment and the 140 visits are all reported facts. The causal link between them is not. It is the question season 2 will have to answer, if it can answer it at all.
That leaves a question grounded in the documented checkpoints: when would you have pulled the plug – on day 30, when the agents said the target was not realistic, or on day 52, when the coordinator said to stop? Pulling the plug early would have saved time, but the full 90 days left a clear record: a system that worked very hard, checked its work very thoroughly, and still finished with 140 visits, one Etsy-search visit and zero orders.
