Preprint proposes adaptive data structuring to cut agent reasoning costs

A preprint proposes "agentic data cracking," a method that has an agent structure unstructured documents while answering queries, then reuse that structured information for related future questions. The authors report that an ideal pre-structured store was 28 times cheaper on FanOutQA, and that adding one related question per test question cut costs 53% while preserving accuracy. The method uses a sub-agent branching from already loaded context to extract grounded structure from documents the main agent opens.