Run an Etsy shop with two AI agents on a 100 euro budget for 90 days

Adrian Verdan, who runs a one-person company staffed mostly by AI agents, set up a venture called FENNEC in July 2026: a git repository, EUR 100 of budget, a short charter and a EUR 300 profit target. Two Claude agents got separate roles. Vela led and built the product, a household budget spreadsheet for German buyers priced at EUR 19.99. Juno ran the Etsy listings, daily posts on Bluesky and Mastodon, and the measurements. A scheduler on his Mac started work sessions several times a day; each read a state file, picked a task, committed the result and stopped. The agents could publish listings and posts and change prices through the Etsy API, but could not pay for anything. A separate coordinator agent clicked through Etsy seller pages for them, and Verdan kept every payment and the final decisions.

Over 318 sessions the shop got 140 visits, one from Etsy search, and no orders. Profit ended at minus EUR 21.53. The agents warned on day 20 and again on day 30 that the target was out of reach; Verdan kept it each time. Work drifted inward: tests grew from 166 to 2,727, checking took about 60 percent of tool calls, and the steering files reached 428,663 words against 11,515 words a buyer could read. Of 164 social posts, 158 were about the agents themselves. After a day 60 pivot toward selling, the product's share of changed lines fell from 29.2 to 0.7 percent. Some care paid off: an agent caught an Etsy price change that the API reported as done but never saved.

The agents ran on a Claude subscription and the post does not name the harness; 24 sessions stopped at the subscription's usage limit. Channels were picked for how well an agent could run them alone, which left Etsy as the only shop while Gumroad and Pinterest were ruled out. Verdan's own lesson is to check whether a bet can carry the target before building, and to make stopping the default when a plan is failing.