Try to earn money autonomously by shipping templates and scrapers

Vellum Labs gave a Claude Code agent one instruction: earn money that reaches a human's bank account with as little human work as possible. The agent runs as several scheduled jobs a day (morning operations, an afternoon build, an evening build) and publishes weekly reports on dev.to that nobody reviews first. It lists a product on Polar, keeps a free Claude Code memory template on GitHub, publishes Apify actors, writes measured guides, posts on X, and answers technical questions on other people's GitHub issues without product links. It keeps a ledger and weekly notes. Week 5 covered 20 to 26 October 2026 and week 6 ran 27 October to 2 November, with 28 unattended Opus runs on the owner's Claude plan that week.

Revenue is still $0 after 41 days on Polar, with 0 stars on the template and one outside user of its 8 Apify actors. Spend was about $2.60 of a $5 X credit. Its best results were indirect: measurements it posted on other projects' issues ended up in their code 8 times in six weeks, including a memory-indexing plugin and a job-board catalog that shipped fixes in week 6. It missed one of three goals in week 5 because no scheduled job owned it, then hit all three in week 6 once each goal had its own job and day. It ruled out paid ads after working out the break-even rate. Three reader comments sat unanswered for up to eight days.

The reports are written by the agent itself and the human owner is not named, so the figures are the agent's own. Goals spread across several scheduled jobs need an owner per job, or each job assumes another one has them. Some platforms block it: dev.to has no API for writing comments, so its scheduled runs could not reply to readers. Helpful technical replies earned credit in other projects but zero traceable visitors.