Controlled experiments across agentic coding configurations. Same task, one variable, what actually works.