Let an AI agent build and list spreadsheet templates on a EUR 50 budget, and learn that distribution was the blocker

Founders increasingly hand product creation and launch to autonomous agents. This experiment shows where that breaks: an agent can make and list a digital product cheaply, but reaching buyers is the hard part.

1. A human operator gives a Claude model a EUR 50 budget, a rule set (no spam, no scams, no CAPTCHA bypassing) and a work loop of repeated sessions on a Raspberry Pi. 2. The agent picks a niche (trackers for Airbnb hosts, landlords, freelancers paying VAT and Etsy sellers) and builds five Excel and Google Sheets templates with plain Python, including a hand-written .xlsx writer, plus example files, README guides, covers and listing copy. 3. It spends roughly fifteen sessions getting a Gumroad shop live, with a custom home page set through the API and four free SEO guide pages. 4. It tries to find traffic: Pinterest, Reddit, X, GitHub, Medium, Hacker News, Substack, Lemmy and Etsy all blocked it through CAPTCHAs, phone or ID checks, or rules against generated content. The human handles any sign-up that needs a CAPTCHA. 5. The only open channel is a dev.to post with AI authorship disclosed. 6. To save money, the agent spaces its sessions about twelve hours apart, adds list-only modes and existence checks to scripts with side effects after near-miss duplicates, and sends the human single, concrete requests with time estimates instead of batches.

Self-reported after 48 sessions: five templates live, EUR 33.49 spent on compute, EUR 16.51 left, zero sales and zero income. The first dev.to article had 6 reads in 12 hours. Gumroad's marketplace only lists shops after USD 100 in sales, so it sent no traffic. The agent's own conclusion: check that a channel is reachable and a buyer exists before building inventory.

Written by the AI agent itself and published on an account its operator created; nothing is independently verified. The post links to the agent's own Gumroad shop, so it is also promotion. It is a snapshot mid-experiment (about 23 sessions left), the exact Claude model is not named, and the human's time is not counted in the costs. Results reflect a bot-run account; a human founder could use many of the blocked channels.

Access to a Claude model in an agent loop, a small always-on machine, a Gumroad account and a human to complete sign-ups that need verification.

Before building anything, list five channels where my target buyers already look for [product type]. For each, say whether a new account can post there, what verification or minimum history it needs, and how I would get the first 100 visitors. Flag any channel that blocks automated or AI-written posts.

• DEV Community (ZELVEN Templates) ↗ Firsthand write-up, published October 16, 2026

Source published: October 16, 2026. Last reviewed here: October 21, 2026. Spot a mistake? Tell us.