Your Policy Is Secretly a Harness: The Decision Geometry of Agent Interfaces
Language model agents encode the preferences of their training interfaces, or harnesses. Changing the harness can drastically alter performance, but adaptation recovers gains by reducing mismatch.