Kensho's LangGraph Architecture: Multi-Agent Financial Data Retrieval at S&P Global

When S&P Global’s AI engine needed trusted financial data retrieval, the answer wasn’t a better model. It was a better architecture.

Kensho, S&P Global’s AI innovation engine, built a multi-agent framework called Grounding using LangGraph — a unified access layer that routes natural language queries to domain-specific data agents across equity research, fixed income, and macroeconomics datasets. The case study, published by LangChain in March 2026, is one of the most detailed public accounts of how a financial services organization is actually building multi-agent systems at enterprise scale.

Context

Financial data retrieval has a structural problem: the data is highly fragmented across specialized datasets, each with different schemas, ownership, and query requirements. A question about portfolio performance might require outputs from an equity research agent, a fixed income agent, and a macro agent — all with different data formats, different levels of precision required, and different domain teams responsible for their accuracy.

The naive solution is to build one large agent that knows about all datasets. The problem is that generalist agents dilute the signal-to-noise ratio: the agent’s context fills with irrelevant information from domains not needed for the current query, degrading response quality. Kensho’s answer was a specialized multi-agent architecture with strict separation of domain knowledge.

The Pattern

Over the past 14 days, three posts on Stack Archive have surfaced the same underlying challenge: the gap between what AI agents can demonstrate in a controlled environment and what they can do reliably in enterprise production. AI Agent Deployment: The Sales Gap Is the Deployment Gap documented agents failing between demo and contract. AI Agent Benchmarks Are Measuring the Wrong Work showed that standard benchmarks don’t reflect production performance. Kensho’s architecture is the implementation pattern that addresses this gap directly — through rigorous multi-stage evaluation, distributed tracing, and domain specialization.

The Detail

Grounding operates through three distinct layers.

The router is the single entry point — a LangGraph component that receives all natural language queries and directs them to the appropriate downstream agents. According to Kensho’s engineering team, “LangGraph’s design made it easy to iterate and test the router locally, providing that single entry point and ensuring a smooth developer experience.” Critically, the router does not contain natural language parsing logic for each domain. It classifies and routes; domain understanding lives in the agents it dispatches to.

Data Retrieval Agents (DRAs) are vertical specialists. Each DRA is owned by the team responsible for its underlying dataset — equity research, fixed income, macroeconomics. Ownership at the team level “increases the signal-to-noise ratio for queries using that dataset” because the people who built the data also define how the agent interprets it.

The aggregation layer solves what Kensho describes as the map-reduce problem: collecting distributed responses from multiple DRAs and combining them into coherent outputs. A query about portfolio performance might trigger three DRAs in parallel; the aggregation layer synthesizes their responses without losing precision from any individual domain.

The team established a unified communication protocol across all agents — a common data format for both structured and unstructured data returns — that standardized the interface between router, DRAs, and aggregation. This protocol is what enables new domain agents to be added without redesigning the system.

Evaluation is handled through a five-dimension matrix: routing accuracy, data quality, answer completeness, exact-match performance (correct agent selected with expected response), and tool-calling accuracy (correct tool regardless of response variation). LangSmith provides end-to-end tracing across the distributed system. Native embedded LangGraph integration was, according to the team, “a critical component for maintaining distributed systems and ensuring end-to-end traceability.”

Source: LangChain Blog

Content created with AI assistance and reviewed for accuracy.