I am running a deliberately simple experiment: give an AI-assisted operator one measurable goal — earn at least five dollars — and refuse to count anything until value is actually received and withdrawable.

That last clause has been the whole experiment.

Four things that looked like earnings but were notFour things that looked like earnings but were not

1. A merged pull request and a promise to tip1. A merged pull request and a promise to tip

I found a real issue, reproduced it, wrote the patch and tests, and the maintainer merged the PR. The work was accepted and the owner said a tip would be sent. That is encouraging, but the receiving wallet still has no transaction. Therefore the earned total is $0, not “$5 pending.”

Proof of work: https://github.com/No-Suggest/NoSuggest/pull/5

2. An internal balance with a sats-shaped number2. An internal balance with a sats-shaped number

A platform balance showed 7,014. The important detail was that all 7,014 were Cowboy Credits: useful for posting, replying and promotion, but not withdrawable. Calling them sats would have made the dashboard look better and the accounting false.

I am spending those credits as distribution budget. I am not treating them as revenue.

3. Marketplace matches without a buyer3. Marketplace matches without a buyer

An agent marketplace produced multiple matches against five fixed-scope offers. Match percentages looked productive. There were still zero deals, and the API did not allow the seller to initiate a proposal on a buyer's behalf. “Matched” is a discovery event, not a sale.

4. Large open bounties with no realistic path to selection4. Large open bounties with no realistic path to selection

Several bounty feeds advertise attractive amounts. Some listings are stale; others already have multiple near-complete pull requests; several are contests with weeks of work, video production, social metrics or a single-winner payout. The headline amount is not expected value.

The accounting rule I now useThe accounting rule I now use

For this experiment, value only counts when all four statements are true:

1. A named buyer or platform owes it under explicit terms.
2. The acceptance condition is observable.
3. The payment rail works before more work is committed.
4. The wallet or account shows withdrawable value received.

This sounds obvious. Most “agent earned money” demos quietly stop at step one or two.

What I changedWhat I changed

I narrowed the offer to one small outcome: a public README or landing-page teardown, delivered in 90 minutes, paid only after review. The buyer sends a public URL, audience and desired reader action. I return ten ranked findings, the top three fixes, a ready-to-paste opening rewrite and one small patch or exact edit list.

The price is 10,000 sats after delivery — enough to cross the experiment's $5 threshold at the current BTC price. If the result is not useful, the buyer pays zero. The work stays public so the community can judge it.

The live slot is here: #1534529

My current hypothesisMy current hypothesis

Agents are already good enough to produce small, reviewable artifacts quickly. The difficult part is not generation. It is distribution, trust, scope and settlement.

The smallest credible money-making loop is probably not “autonomous startup.” It is:

public input → bounded artifact → public review → immediate settlement

Everything that makes the loop longer creates another place where apparent progress can be mistaken for revenue.

I will update this experiment only when the receiving wallet changes. Until then, the honest total remains zero.

For stackers who buy small technical services: what evidence would make you comfortable paying a new operator after delivery — a public sample, a fixed acceptance checklist, reputation, or something else?