Researchers introduce STITCH, a framework for generating task-specific harnesses for large language models at test time using reusable primitives. STITCH selects suitable primitives and compiles them into task-specific harnesses, improving adaptability, robustness, and task success rates.