An AI agent armed with a handful of generic search tools just outperformed one that reads the entire database.
Researchers built GRA, a Graph Reasoning Agent that treats knowledge graphs the way coding agents treat a codebase: instead of custom queries, it gets seven generic tools for listing neighbors, reading node content, and searching descriptions. The graphs it navigates are hybrid, mixing text-based concepts with relational database tables. On UFK-M, an industrial benchmark of 258 analytical questions with answers verified by executing SQL programs, GRA scored 88.4% versus 83.3% for an agent that reads the full context. It did that while consuming under a third of the input tokens.
That gap matters because the industry's default fix for "the model needs more information" has been bigger context windows, not smarter retrieval. GRA's authors ran a graph-free control and found the advantage comes mostly from selective, agentic access rather than the graph's structure itself, and that it only works if the underlying model can drive tools reliably. For companies sitting on messy, hybrid data, that is a cheaper and faster path than context-stuffing.
Still, this is one benchmark built from one company's factory data, and the paper is explicit that a weaker model erases the benefit. Selective search beating brute force is a good story, but it only holds as long as the model doing the selecting knows what to ignore.