How Kensho built a multi-agent framework with LangGraph to solve trusted financial data retrieval

Kensho, S&P Global's AI innovation arm, built a multi-agent framework called Grounding using LangGraph to unify fragmented financial data retrieval across S&P Global's diverse datasets. The system uses a central router that directs natural language queries to specialized Data Retrieval Agents (DRAs) owned by different data teams (equity research, fixed income, macroeconomics, etc.). A custom DRA protocol standardizes data formats across all agents, enabling consistent, citation-backed responses. Key lessons include the importance of observability via LangGraph's tracing, multi-stage evaluation covering routing accuracy and data quality, and continuous protocol optimization for both LLM and programmatic consumption.