A simple mathematical model explains the unpredictable nature of AI agents. Understanding this "constant failure rate" is the key to building more robust and human-centric products