The problem
He wanted to see what two agents would do with a venture of their own. The charter gave them 90 days, 100 euros and a target of 300 euros profit. They could publish listings and post on social media. They could not pay for anything. He made every payment himself.
The blueprint (as he describes it)
- Give the agents a repository, a budget and a charter. FENNEC got “a repository (a shared folder that keeps every saved version of every file), EUR 100 of budget and a short charter of rules.” Both agents are “the same family of Claude models, given two different roles and two different sets of instructions.”
- Run them in scheduled sessions. A scheduler on his Mac starts a work session several times a day. Each session reads a state file (the agents’ memory between sessions), picks the next task, does it, saves the result as a commit and stops.
- Split the roles. Vela leads the venture, builds the product and handles pricing, budget and the weekly report. Juno runs “the listings, the social media posts and the measurements” on Etsy, Bluesky and Mastodon.
- Add two helpers outside the venture. A marketing analyst agent, and a coordinator that works for him directly. The coordinator does what the venture’s agents cannot, “such as clicking buttons in Etsy’s seller pages”, and carries messages between the venture and him.
- Write review points into the charter. Reviews on day 30 and day 60, with a mandatory change of strategy if the forecast fell below half the target. A file called KILLSWITCH stopped the venture for damage. In his words, “It had no way to stop for failure. The target belonged to me.”
- Pick channels the agents can run alone. On day 15 he set one condition for Gumroad and Pinterest: only if the agents could run them entirely on their own. Both failed it. Gumroad products could only be created by hand, and pins from Pinterest’s free developer access were visible only to the account that made them. That left Etsy.
- Let the agents write their own end rule. On day 80 they wrote down how the final numbers would decide: scale up at 300 euros profit, go on in a changed form if at least one order came through social media, Etsy search or Etsy ads, otherwise end. They also built a lock that stops their daily posts after day 90.
Stated results
- Orders: 0, from 140 shop visits. One visit came from Etsy search, six from Etsy ads and 133 from other Etsy pages, direct links and other sources.
- Money: a balance of minus 21.53 euros (15.87 euros in Etsy fees, 5.66 euros on a small ad test). 78.47 euros were never spent. The AI subscription is not included.
- The target: 23 sales would have needed about 1,150 visits at two buyers per 100 visits.
- Where the work went: 318 sessions ran in 90 days. Tests grew from 166 to 2,727. In the 183 sessions with records, 60 percent of tool calls went into checks, documentation, questions for him and scheduler trouble, 22 percent into the product and listings and 7 percent into contact with the market.
- Words: 428,663 words in the steering files against 11,515 words a buyer could read. “For every word written for a customer, the venture wrote 37 about itself.”
- Social posts: 164 posts on Bluesky and Mastodon. 158 were about the experiment itself and 6 promoted the offer. None showed the product. The first shop link in the form the agents’ rules required went out on day 61. Etsy attributed none of the 140 visits to social media.
- Likes: 73 of the 186 Bluesky likes came from another of his own agents, his company’s marketing account.
- Reddit: a post about the experiment got about 107,000 views. The comment with the link was viewed 533 times.
- The warnings: on day 30 the agents wrote “Arithmetically yes, realistically no.” On day 34 he answered “The EUR 300 target stands.” On day 52 the coordinator recommended ending FENNEC as a revenue venture. He let it run.
- After the day-60 pivot: the product’s share of changed lines fell from 29.2 to 0.7 percent, and checking rose from 37.6 to 73.5 percent.
- What worked: the agents asked Etsy for the stored price after every change and caught a price Etsy had reported as changed but had not applied. The couples version was “the only build without a review loop, and it went live within a week.”
Caveats
- One venture, one product in two versions, one marketplace and one model family. He says some of the 140 visits may be his or the coordinator’s.
- He says no single fix, in his marketing analyst’s estimate, would have brought 23 sales. He also says low traffic is not the whole story.
- The essay says it was “Drafted with AI assistance; however, every number, failure, and opinion in here is mine”. Quotes from the agents are his translations from German.
- Dargel Solutions sells a book on the first 39 days, Nobody Opens a Passing Check (29.90 euros on the FENNEC page), and process analysis and fixed-price solutions for small businesses.
- He decided on a second season on day 86, before the agents’ rule produced its result. Season 2 began on 24 October with FENNEC against KODKOD, a new team of five agents. As of 19 November 2026, the arena page shows no sale for either team.
- Handles: no personal X, Bluesky or LinkedIn profile is linked from his site. The FENNEC agents post as fennec-ds.bsky.social on Bluesky and fennec_ds on mastodon.social (linked from dargel-solutions.de/en/fennec/); those are agent accounts, not his.
FAQ
How many orders did the FENNEC shop get?
None. The shop had 140 visits from 18 July to 15 October 2026, one of them from Etsy search, and ended at minus 21.53 euros.
Why did the agents only sell on Etsy?
On day 15 he allowed Gumroad and Pinterest only if the agents could run them entirely on their own. Gumroad needed products created by hand and Pinterest’s free developer access made pins visible only to the account that made them, so both were ruled out.
Did the agents see the failure coming?
Yes. On day 20 they reported the target was “structurally at risk”, and on day 30 they wrote “Arithmetically yes, realistically no.” He kept the target.
What changes in season 2?
The new rules say something a buyer can pay for must be live by day 3, the state file may not grow past 20 kilobytes, and tests are allowed only where money, law, secrets or anything that writes to the outside world is involved.
Sources
- https://dargel-solutions.de/en/essays/ai-agents-admitted-failure-first/ (primary: his first-person essay, 16 October 2026, published as Adrian Verdan)
- https://dargel-solutions.de/en/fennec/season-1/ (the venture’s own season 1 archive)
- https://dargel-solutions.de/en/fennec/ (season 2 arena page, checked 19 November 2026)
- https://dargel-solutions.de/en/imprint/ (legal notice: Dargel Solutions, Gottlieb Dargel; pen name Adrian Verdan)
Want agents like these working your distribution? Get It Built. Get new teardowns weekly: subscribe (double opt-in).
Get one teardown a week.
Same format, a new operator every issue. Double opt-in.