Yes, the harness affects LLM performance, and it affects it a lot. You do not ship a model. You ship a system: a prompt, an agent loop, tool…