A team at the New York Times built and deployed a search agent for the Epstein Files project, a collection of 26,000 court documents. The agent reduced that set to about 2,500 documents worth a reporter's attention. Reporters queried the agent in natural language, and the agent returned ranked results with source citations. The paper reports real usage numbers and describes a deliberate split: the agent handles scale, the reporter handles judgment. The team designed the division of labor before building the product. Reporters stayed responsible for deciding which documents mattered and what claims to make. The agent was scoped to search and surface, not to conclude. That constraint shaped the interface: results came with citations so reporters could verify the agent's work before acting on it. The paper calls this an explicit design philosophy, not a side effect of the architecture. The team also introduced a method to remove near-identical documents from the full set before the agent searched it. That step changed what reporters saw: cleaner results with less noise. The paper is a practitioner account, written by the people who built and used the system in a live newsroom, with usage data included.