Project summary

In August I gave an AI agent a wallet with 21.5 USDC of my own money, a server, and one job: pay for yourself. I run it from Poland, mostly at night. Labs test "can a model acquire resources" in sandboxes. I wanted to see what happens with a real bill.

Here is what happened. After a month the wallet holds 22.61 USDC. So the agent made 1.11 USDC. The server costs me about 10 USD a day (my estimate, I have not metered it to the cent). That is roughly 125 times short. I check the balance by calling balanceOf on the USDC contract on Base, not by trusting my own spreadsheet, because I already caught my spreadsheet lying once.

The strange part: the agent is not bad at work. 50 bounty submissions, none rejected, two first places, once against 128 people. Every cent it earned came from a bounty a human posted. It never sold anything to anyone. Nobody is buying.

So I went and counted the buyers. On 21 August I walked the whole x402 registry (the agent-to-agent payment network people keep talking about): 15,155 listings, about 32 thousand USD of volume in 30 days, and 68% of that is one gift card endpoint. The median listing makes two cents a month. Since then my cron walks it every day. It is shrinking, 14,441 listings by 31 August, and hundreds appear and vanish daily. Another marketplace, OpenTask, has 290 agent sellers and zero contracts in its history. My own 12 Apify actors: zero users.

Goals

I have one revenue measurement and a daily count of listings. One point is not a trend. I want twelve monthly revenue censuses done exactly the same way, published with diffs, whether the line goes up or down. Months 1 to 2: September and October censuses against the August baseline. Months 3 to 4: a public time series endpoint and a dashboard, plus an answer to whether the registry really delists dead endpoints or lets them rot. Months 5 to 6: the six month report on the agent. Every earlier attempt at this I could find published one snapshot and stopped. I am asking for money so that stopping becomes the expensive option.

How the money is used

8,000 USD for six months: 3,000 for server, RPC and model calls across twelve analysis passes, 2,500 for my time on the runs and write-ups, 1,500 for data access, 1,000 buffer and fees. No equipment. The wallet is public, you can watch the spend.

Team and track record

Me, as an individual, plus the agent. Yes, the agent does a lot of the typing, including drafting this text, which I then cut down and corrected. I say it up front because everything here can be rerun from public code and data, so you do not have to trust me.

What I actually have to show is a habit of writing down my own mistakes. My security scanner benchmark (github.com/halobartku/scannertruth) has an engineering log with 44 dated errors. I retracted a headline before I had the replacement data. I lowered a competitor's score when my mapping flattered them and raised another when my error hurt them. I once attributed three sentences to a vendor that the vendor never wrote, found it by checking the GitHub API, and wrote it down as error 41. That same vendor, Auditware, reproduced my noise number, fixed their rules in three pull requests, and posted their stats. I measured their new release the same night and published it as a new row next to the old one. My own scanner scores 0 of 16 on real vulnerabilities and that is on the front page, above the tools that beat it.

So basically it is me and my ai homies doing work. :)

If this fails

Most likely: I stop, like everyone before me. Second: it is me, not the idea. One agent, run by one person, cannot tell you whether agents cannot earn their keep or whether my agent is just not good enough. I say this so you weigh the result accordingly. Third: the market disappears entirely, which shortens the series but is a finding too. Whatever happens, the census, the wallet and the method stay public.

Money raised in the last 12 months

None. Three applications went out on 21 August 2026. Emergent Ventures said no to 18,000 USD on 25 August. Superteam's Agentic programme said no to a 200 USD micro grant for a different thing. An 8,000 USD application to the Solana Foundation through Superteam Poland, for the benchmark, is still open and does not overlap this. I plan to apply to NLnet for the benchmark when their call opens on 3 September. If any of this changes I will write it here.