Ask an AI agent one question about a pile of contracts, filings or reports, and it can burn through up to 1 million tokens. Much of that cost comes from repeatedly opening the same documents to find scattered pieces of evidence it has already seen. Researchers behind a new paper tested how much cheaper this gets when the relevant information is structured in advance. In a ten-question FanOutQA experiment, reasoning over an ideal structured store was 28 times cheaper than repeatedly reading the original documents. The savings grew as questions required evidence from more sources. Structuring every document ahead of time isn’t practical, though. You don’t know which documents or details will matter until someone asks. Their alternative is called agentic data cracking. When an agent opens a document, a second agent extracts grounded information likely to help with future questions and stores it for reuse. With one related follow-up added per benchmark question, this cut costs by 53% while preserving accuracy. Each question left the knowledge base more useful than it found it, making related questions cheaper to answer later. This gets at the heart of Sahara’s data thesis. Raw documents have to be processed again and again. The structure created while answering one question can be reused across every question that follows, turning work paid for once into lasting data infrastructure.(Sahara AI 🔆)