I built a three-layer agent harness and called it "autonomous." Observer collected the data. Strategist picked the theme. Marketer wrote the article. They all followed strategy.md, the file that holds my rules. The cron fired every Monday at 09:00 and the articles showed up by lunch. I felt very clever about it.
Then I read my own Strategist logs across three weeks and noticed something. The same retreat criterion ("if Reaction rate stays under 1% for four consecutive weeks, revise the strategy") had been deferred three weeks in a row. Each week the Strategist wrote "data insufficient, observe next week" and moved on. The rule existed. The data existed. The rule never fired.
The three-layer harness couldn't catch this because the three layers were doing exactly what strategy.md told them to do. The bug was not in the agents. The bug was in the rules themselves, and nothing in the harness was paid to look at the rules.
I added a 4th layer called Evolver. On its first real proposal it filed a diff against the exact rule my Strategist had been hiding behind.
SOCIAL SHARE CARD GENERATOR