The paper identifies task decomposition — not retrieval — as the binding constraint in multi-skill agent planning, and SAD's single-iteration fix raises decomposition accuracy by over 32 percentage points, directly improving how reliably agents can assemble executable plans from large real-world skill libraries.