Everyone has an opinion about which model is best. Far fewer people can show, with production data, which levers actually change an agent's output quality. Drawing on OpenTelemetry instrumentation of real Claude Code sessions and on enterprise rollouts where consistency matters more than demos, this talk works through the levers one at a time: context engineering, reasoning effort, permissions and hooks as behavioral rails, eval loops, and the narrower set of cases where model choice genuinely matters. For each lever you get what it costs, what it moves, and how to measure the difference in your own environment. Claude Code is the running example; the levers apply to any serious coding agent. The field changes monthly, so the content will be current as of the conference, not as of this abstract.