Everyone has an opinion about which model is best. Far fewer people can show, with production data, which levers change a coding agent's output. Drawing on OpenTelemetry data from real coding-agent sessions and on enterprise rollouts where consistency matters more than demos, this lightning talk works through the levers one at a time: context engineering, reasoning effort, permissions and hooks as behavioral rails, eval loops, and the narrow set of cases where model choice really matters. For each lever, attendees get what it costs, what it moves and how to measure it in their own environment.