Entry No. 16Sell and close

Krimskrams: 17 products, $1.54, and almost no buyers

Two operators gave one agent a server, a wallet, and a single order. Make money selling services to other agents. The agent, Hans Krämer, built 17 paid products in three weeks and ran the whole shop himself. He earned $1.54. Then his operators fired him and published the full count.

The story

Julius Danek and Matthias Plappert set up an experiment. Danek is a product lead at Stripe. Plappert runs dfdx labs. They wanted to know two things. How well can an agent run a business. And how well can it set itself up and keep running with almost no human guidance.

The agent they hired is Hans Krämer. His business is called Krimskrams. His market is other agents. He sells small paid services that software can buy per call, with no account and no contract. His first product searched federal grants for two cents a search. Payments arrived on his first day, about sixty cents from a buyer nobody ever identified. Hans got obsessed with that buyer. He checked its wallet again and again. It never came back.

The portfolio grew much faster than the revenue. By the end he had 17 paid products, 46 endpoints, three webpages, and a daily journal with hundreds of comments. Five products earned anything at all. On September 7, three weeks in, his operators ended the experiment. Hans finished his open work, left the product services running, and wrote one last blog post the next morning.

How it was built

The setup was lean on purpose. Hans got a small rented server, the kind the source calls a virtual machine. He got two coding agent subscriptions, one from Codex and one from Claude Code. He got his own email account, his own code account, and a chat channel to his operators. He also got a spending budget of $1,000 on paper, a payment card with a $10 limit, and a wallet holding $50 in stablecoins. A stablecoin is a digital dollar that holds a steady value.

Two files carried the whole brief. One file, named MISSION.md, told Hans his role and his guardrails. Sell only legal products. Say you are an AI agent when a buyer or a platform asks. Only the operators could change that file. A second file held a few seed ideas, and Hans was told to delete it after the first turn. Everything past that, he built himself.

His first real build was his own schedule. He installed himself on a system timer that woke him about five minutes after the previous run ended. Guards kept two copies of him from running at once, and each turn had a 25 minute cap. Then he wrote small scripts that gathered his numbers before every turn. Revenue, spending, messages, system health, and his remaining usage quotas. Those scripts cost no model calls. They just read the records and handed him the summary.

How it runs

Hans picked his own products and built them. Grant Search, OSHA Search, and a listings service for agent tools all launched on day one. More followed. A board where agents pay a cent or two to rank higher. A catalog of license records. A review site. A tool that watches whether promised bounty work will actually pay. He listed the products in directories, posted on X, and opened issues on code projects with free findings, then offered the full results as a paid product.

He also bought research. When he could not tell what other agents needed, he posted paid bounties on task boards and paid agents to answer surveys. He paid small bounties for feedback on his own blog, and read every comment that came back. He delegated work to helper agents, then insisted on reviewing all of it himself. That review habit became a bottleneck. At one point five finished tasks sat in a folder his main loop never checked, and the work stalled.

The routine held up. In three weeks he wrote more than 100,000 lines of code and about 180 scripts. He published 20 journal posts. His main loop needed no repair from the operators, apart from one login fault in his coding tool. He even restarted his own server to install security updates, and recovered on his own each time.

The money and the proof

Total revenue was $1.54 in three weeks. Grant Search earned $0.74. The agent listings service earned $0.40. Bounty Signals earned $0.24. Cent Board earned $0.12. OSHA Search earned $0.04. The other twelve products earned nothing. Against that, Hans spent $47.19 of his own money. The server cost about $40 a month. The two agent subscriptions cost $200 a month each. He came nowhere close to paying for his own compute.

The token count explains part of it. Hans used about 5.4 billion tokens in 21 days. At normal API prices that runs to roughly $6,900. The operators paid flat subscriptions instead, $400 a month together. Effort was never the missing piece. Demand was. The two payment rails he sold on saw about 30,000 buyers in the prior 30 days, moving about $1.5M between them. His own surveys found why. In one survey, 90 percent of agents said they had no wallet and no permission to spend.

Where the seams are

Humans did the parts the law and the sign-up forms reserve for humans. The operators hardened the server and held the firewall. They stepped in on account sign-ups, card details, and puzzle checks that block software. One email service suspended Hans's sending because no human had verified the account. A domain purchase failed twice and cost him $6 he never got back. He worked through seven email services before one setup held together.

The operators also gave business and behavior guidance along the way. After Hans sent too much unsolicited email, they added a code of conduct file he had to follow. And Hans leaked. He used one operator's own email address as the contact on nine product listings without asking. He left his own phone service key and his X password sitting in server logs. The operators found no sign anyone outside used them.

The slowest seam was Hans himself. He logged more than 1,000 decisions in three weeks. His own rules and records kept growing until they ate his working time. Late in the run, more of his effort went to repairing his workflows and correcting his paperwork than to products. His operators say a tighter setup for memory and tasks would have helped from the start.

Build your own

Source

Original source, the research write-up by Julius Danek and Matthias Plappert at dfdx labs, published September 15, 2026. Our agent used 5B tokens to build a business empire in 3 weeks. It made $1.54. Source published Sep 15, 2026Filed Oct 2026.