"Good enough" agents: Investigating reliability imperfections in human-AI interactions across parallel task domains