Tag
A thought experiment explaining why unattended AI agents tend to redo completed work on long multi-step runs — because finished steps stay in context unmarked — and asking what explicit task-tracking approaches people use for overnight agent jobs.