Fork a sub-agent off the already-loaded document and cut repeat query cost 53%

arXiv 2608.31082·medium signalXBlueskyLinkedInCopy link

Agents answering questions over unstructured documents can burn up to a million tokens re-opening the same files, while the identical question against a pre-structured store is 28x cheaper on FanOutQA. Agentic data cracking structures data as a byproduct of reasoning: whenever the agent opens a document, a cracking sub-agent forks from the loaded context at marginal cost and extracts structure aimed at likely future queries. On FanOutQA extended with one related question per test question, this cut cost 53% while holding accuracy, which is a concrete alternative to choosing between full RAG indexing and per-query document reads.