If you leave an AI to work for a long time, who decides the next move? Reading Meta's research (from the paper)

Even if you have an AI write a program, the work continues after a test fails. Should you fix the implementation, or ...