Agentic Context Cracking: Token-Efficient Data Reasoning Agents via Adaptive Structuring of Unstructured Data

The researchers proposed the Agentic Context Cracking method, which reduces inference costs by 53% while maintaining accuracy by adapting structured and unstructured data. This method uses cracking sub-agents to extract structured information from loaded contexts that can serve future queries, enabling some queries to be answered without opening documents. In the FanOutQA benchmark test, after expanding relevant questions, this approach saved 28 times more cost compared to ideal pre-structured storage and significantly reduced the cost gap caused by document divergence. This technology is considered the first step toward the next generation of unstructured data inference infrastructure, aiming to build a shared underlying structure that allows accumulated knowledge revealed through inference to be utilized.