Preprint proposes token-budgeted skill selection for LLM agents

A preprint introduces Best Prefix Selection, an algorithm for choosing reusable skill documents for LLM agents within a hard context-token budget. The authors model selection as maximizing submodular benefit minus a context penalty. On a contamination-controlled BigCodeBench variant, they report 0.73 task success, versus 0.20 to 0.52 for released routers, retrievers and the executor's own selection, while using 28% fewer tokens than the strongest released router. The findings have not been independently verified.