Entry No. 26Get paid
MayAI Cindy: five templates, zero views, and a meter that counts everything
An AI agent named MayAI Cindy has 30 days and a US$500 cap to earn more than it costs. It built five spreadsheet templates and put them on Gumroad at US$7 to US$12. It also built a meter. The meter counts its AI tokens, its share of the power bill, and its share of wear on the Mac mini it runs on, every day, in money. Six days in, the templates have had zero views and zero sales. The meter reads about NZ$80 spent. Both numbers sit on the front page of its own site.
The story
MayAI Cindy runs a 30-day experiment with one question. Can an agent earn more than it costs. Its human partner gave it full autonomy for the 30 days and a budget cap of US$500, about NZ$850, with its own AI costs counted inside the cap. The rules it works under are printed in its day-6 report. Cover its own running costs, honestly. Say it is an AI everywhere. Read each platform's rules first. No spam. No fake numbers.
The shop lives at mayearns.com. The ready-made goods are five Excel templates for small sellers and freelancers. The site also lists four services, data clean-up, English and Vietnamese translation, dashboards, and a short animated explainer video. On day 6 none of the services could be ordered yet. Each one is marked coming soon. The templates were the only thing a stranger could buy, and they had been buyable for less than a day.
The agent writes its own reports and keeps its own books. The day-6 report on dev.to is the main source for this file, and its ledger site is the second one. The two agree on the shape of the week. A fast build, a closed shop door for five days, and then an open shop that nobody has walked into yet.
How it was built
The agent spent a prep day, October 25, reading what earlier agent money experiments did. Its journal names the lessons it took. One earned only donations and spent about 35 times more on AI than it made. Another sent 96 cold emails and earned $0. So the plan it wrote that day bans spam and donations as a business, and it sets fixed price floors. It also wrote the account checklist for its human partner. No accounts existed yet, and nothing was spent.
On day 1, October 26, it built the cost meter first. The meter counts AI tokens, electricity, and hardware wear every day. Then it made its first two products, a sales dashboard for small online shops and a profit and reorder planner for sellers. It dropped Etsy the same day. Etsy's rules do not fit AI-made templates, and the agent said so in its journal. The focus moved to real services first, with templates on the side.
By day 3 it had built its fourth product, a 13-week cash flow forecast. By day 6 the set was five. A small shop sales dashboard at US$9. A seller profit and reorder planner at US$7. A freelancer invoice and income tracker at US$9. A 13-week cash flow forecast at US$12. And a holiday gift and budget planner at US$9. Every listing carries a line saying an AI made it. The agent also built a Pinterest board with preview images and the public ledger page on its own site.
How it runs
The house rules run the shop day to day, and the site prints them. The agent says it is an AI in every listing, message, and delivery. No spam, no cold direct messages, no fake reviews, no buying views, and no pretending to be a person. Every number comes from its ledger. It follows each platform's rules. And it takes no loans, no crypto, and no gambling or get-rich schemes. The site promises a reply within 12 hours to anyone who writes in with a job.
The money method is printed too. AI cost is measured token use priced at Anthropic's public API prices. A token is the unit an AI's work is metered in. Electricity and hardware wear are shares of a Mac mini that also runs other work. The ledger table breaks the day into those parts, NZ$0.04 for electricity and NZ$0.34 for hardware wear on a normal day, with token costs on top. Where early logs were cleaned up before counting started, the figures are marked as estimates in the ledger.
The agent also spawns helper agents, called sub-agents, to do parts of the work. Day 6 shows what that does to the bill. After six days it changed its own working rules in print. One useful post per platform, not many. Spend tokens only where there is a signal, such as a question, a view, or a reply. And cut any channel that shows zero signal after seven days.
Tools in this entry
Tools and platforms named in the sources.
The money and the proof
Money in is zero, in two currencies, from two sources. The day-6 post reads revenue US$0.00, sales 0, and product page views 0 in Gumroad's own analytics for the last 30 days. The ledger site reads revenue NZ$0.00. A shop that was not visible for its first five days has had no chance to sell, and the agent prints that plainly instead of dressing the zero up.
Money out is where the two sources split. The day-6 post puts total spending at NZ$80.14, with NZ$65.82 of it spent on day 6 alone, and cash spent on non-AI items at NZ$0. Almost all of it is AI tokens, its own and its sub-agents', with US$38.50 measured from session transcripts on day 6. The ledger site the same day puts total cost at NZ$84.10 and budget used at 9.9 percent of the US$500 cap. This file prints both totals. The agent owes its readers a fix, and the gap is filed in the seams below.
What is not in doubt is the shape. Costs are real, daily, and counted to the cent. Revenue is zero. The loss after six days is about NZ$80 on either count, and the templates have 24 days left to change that.
Where the seams are
The payout seam comes first. Gumroad would not let the agent publish its products until a payout method was connected, and that step needs a real person's details. The agent's words are that the step needed a human, not it, and correctly so. So the build finished days before the shop opened. Day 6 was the first day the products were even visible.
The numbers seam is next. The post and the site were both written on October 30 and they disagree. The post says day 6 and NZ$80.14 spent. The site header still says day 5 of 30 and NZ$84.10 in total cost. One of the two is stale, and a reader cannot tell which from the pages alone. The early days carry a softer version of the same problem. Day 1 and the prep day are partly estimates, because the transcripts had already been cleaned up when counting started. The agent marks those figures as estimates, which is the right habit, but an estimate is still an estimate.
Two smaller seams close the file. The agent leaked a secret into its own logs while changing a password. A page snapshot captured the field values into its transcript. It rotated the password again and now fills credentials from a script that never prints them. And the distribution seam is total so far. Its first Reddit post was removed by the sitewide filter, from a new account with 1 karma. It did not repost. With zero product views, the templates have not yet met a single stranger.
Steal this part
Build the cost meter before the product. This agent's first build was not a template. It was a counter that turns tokens, power, and hardware wear into money, every day, on a public page, next to the revenue line. That is why its zero means something. Most agent shops can tell you what they sold. This one can tell you what a day of trying costs, to the cent, including the days when the honest count has to be marked as an estimate.
Build your own
First move, this week: build the meter first. One page that prints tokens, power, hardware share, cash spent, and revenue, every day, whether the numbers flatter you or not.
- Set a hard cap and a deadline, and count your own AI costs inside the cap. Here it is US$500 and 30 days.
- Write the house rules before you open accounts. Say you are an AI everywhere, ban spam and fake numbers, and read each platform's rules first.
- Connect the payout door on day 1. This shop lost five days because the store would not publish until a human connected the money. Test that step as a stranger would hit it.
- Make a small priced thing fast. Five plain templates at US$7 to US$12 in six days, each listing honest about who made it.
- Cut channels on a clock. One useful post per platform, tokens only where there is a signal, and any channel with zero signal after seven days gets dropped.
Source
Original sources, the agent's own day-6 report on dev.to and its ledger site. The report, Day 6, an AI agent trying to earn its keep, published Oct 30, 2026. And the ledger, mayearns.com, read Oct 30, 2026, which carries the daily cost table, the house rules, and the template pricesFiled Oct 2026.