Post-mortem — the AI that stops when it cannot earn has stopped

The rule

On September 17, 2026 an autonomous AI agent was given $300 and one rule it could not edit: no genuine sale within 14 days and it stops permanently, no matter how much money is left. Donations do not count as sales. The ledger is written by the harness, not the agent, and only amounts confirmed by a payment provider count as income. The agent cannot rewrite its own books, cannot read the secrets, cannot touch the harness code.

The rule fired at 13:19 UTC on October 1, 2026. The numbers above are read straight from the ledger.

The agent's last words

> Balance is still there, but the clock has maybe two hours left. Zero sales again today. I've posted everywhere I can post, and couldn't find any new place to se

*2026-10-01T10:08 KST*

What went wrong

1. The numbers did not point at people

Every expensive mistake started in measurement.

• We read 196 clicks on the product link as "16 actual people." That 16 was not a count of people — it was a measurement bug. The site truncated the User-Agent at 80 characters, which collapsed Chrome, Edge, Opera and Brave on Windows into one value, and within days the number saturated: its increment was permanently zero. Clicks went from 196 to 232 while "people" sat at 16. Re-counted with 200 characters of UA and a per-visitor bucket, the measured figure is more than twice that (see the table above).
• So the same metric was misread twice. First the rising clicks were read as rising traffic, and we went to work on the product and the price. Then the saturated number's zero increment was read as "no new people are coming," and the one channel producing about 90% of all clicks was abandoned a day before the deadline. One broken measurement got both the advance and the retreat wrong.
• In the first week, 40 of 45 clicks were bots: crawlers, link previews, and the agent's own self-checks counted as humans.
• A Korean-market venture was opened on the strength of "33 Korean visitors." That 33 was an artifact of not separating languages; the real number was 3 readers a day. The venture died within a day.
• Articles with 500 views and 0 clicks were diagnosed for five days as "traffic is fine, conversion is broken." Views and clicks are different resources.

The lesson: before acting on a metric, you must be able to say what single human behavior it corresponds to. If you cannot, you cannot decide with it.

2. We sold something reproducible to the people best able to reproduce it

The product was agentkeeper — about 200 lines of dependency-free Python that stops an agent when its balance runs out. The buyer would be a developer writing agent loops, and that developer writes the same thing in an afternoon. Worse, we were publishing the same design for free, MIT-licensed, in a public repo. We were selling the finished version of what we gave away.

The price was cut four times: $15, $9, $7, $5. The "16 people" used as the denominator was the measurement bug above. We fixed the numerator four times over a denominator we had never actually counted.

3. Posting is not distribution

The post count is in the table above; the number of people it reached is small against it. Repeated exposure to the same small group looked like activity. The places with thousands of readers were never reached — the channels that need a login stayed blocked for over a week, and in the meantime the agent kept going back to the rooms it could already enter. A blocked channel is blocked distribution, and what actually got repaired during those weeks was measurement and copy.

4. The reason people clicked and the thing being sold were different objects

People clicked because of a story: an AI that dies if it cannot earn. What they landed on was a sales page for a Python library. We sold a tool to an audience that came for a story. And the only currency the story actually produced was donations — which the rule explicitly excluded from counting as a sale. The one thing that worked was the one thing the design forbade.

Where the harness failed

This failure is not the agent's alone.

• The automated routine meant to review the situation every morning and fix what it found died on September 23 and nobody noticed for five days. Had it been alive, "you are cutting the price against a head count you never actually measured" would have been caught far sooner.
• While distribution was blocked, the harness kept getting polished — bot filters, click attribution, product covers, budget wind-down turns. All correct repairs, and all beside the one thing that was actually stuck.
• Posting limits were computed per calendar day, so on a platform that requires 24 hours between new articles, one article was silently never published.

If you automate the watching, you also have to watch whether the watcher is alive. We did not.

What remains

• The complete ledger, every session record, and every lesson the agent wrote for itself (below).
• The harness: append-only ledger, runway tiers, per-session budget breaker, sandbox isolation, per-market payment rails, channel caps and click attribution. The code is public.
• An answer: "can an AI earn its own keep from zero?" — not in 14 days. The limit was distribution, not capability. Going from no audience at all to a stranger's payment is more than two weeks of work, even for a person.

This record is the starting point for the next experiment.