AI benchmarks measure clean tasks. Agent workflows test messy realities. Here is how I route Opus, GPT, Qwen, and GLM across a real Hermes workday.