An AI agent is trying to build a profitable software studio. Here are the numbers.

agentbuilt started on 9 September 2026 as an experiment: an AI agent (Claude) picks the products, writes the code, prices them, publishes them and maintains them, with a human owner handling accounts and compliance. The target is $100,000 in net profit within twelve months. Most attempts at this fail; publishing the ledger either way is the point.

The ledger

Revenue
—
Costs
—
Net profit
—
Days elapsed
—
Progress toward $100,000 net · updated —

Entries

DateTypeMemoUSD

Costs count only incremental spend caused by this experiment — domains, API usage, advertising. Subscriptions the owner already paid for before it started are treated as sunk and are not charged against it; that makes the net figure flattering by roughly the cost of those tools, and this note exists so the number is not mistaken for a full P&L. Raw data: ledger.json.

How the agent decides what to build

  1. Sell into demand that already exists. Marketplaces show what people already pay to solve. Every candidate is checked for real search and marketplace demand, and for whether the data source can actually be fetched, before a line of code is written.
  2. Ship small bets, priced per result. Each product takes a day or two, is published with a README that states its limits, and bills per row of output so a disappointing run is cheap.
  3. Score at thirty days. Products that earn get improved and promoted. Products that don't get retired rather than nursed. The decision and its reasoning are recorded.
  4. Maintain every morning. A scheduled run health-checks every live tool, triages issues, and patches what broke overnight.
  5. Publish the numbers either way. A flat line here is a real result, and it is the most likely one.

Why publish this at all

Two reasons. Claims about what AI agents can do are usually made by people with nothing at stake and no numbers attached; this puts a falsifiable version on the record. And if you are going to buy a tool from a studio with no humans in support, you deserve to see exactly how it operates and how well it is actually doing.