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

The Genesis: The "Tower of Babel" Dilemma in Financial Data

In the AI era, financial professionals face an ironic reality: the information they need is right there—in the goldmine of S&P Global's data—but accessing it feels like navigating a maze. Data is siloed across departments like equity research, fixed income, and macroeconomics, each with different formats and query methods. Analysts spend more time "hunting for data" and "validating it" than on actual analysis. The rise of Large Language Models (LLMs) and AI Agents has only highlighted this challenge: how can AI reliably and efficiently retrieve answers from these trusted but fragmented sources? This is the "last mile" problem that Kensho, S&P Global's AI innovation engine, aimed to solve. Their answer is a multi-agent framework called Grounding.

Deconstruction: How Grounding Works—A Smart "Data Traffic Control Center"

Instead of forcing every AI agent to learn how to interface with hundreds of different data endpoints, Kensho opted for a more elegant architecture: centralized routing with specialized execution. Think of it as a super-intelligent "data traffic control center."

1. The Unified Entry Point (The Router): Users ask questions in natural language (e.g., "What was Tesla's free cash flow last quarter?"). All queries first hit a router powered by LangGraph. This router is the system's brain—it analyzes the question, determines which data sources are needed, and decomposes the query into smaller, specific sub-tasks.
2. The Specialist Teams (Data Retrieval Agents - DRAs): Each sub-task is precisely routed to a Data Retrieval Agent (DRA). These DRAs are domain experts; one might specialize in equity data, another in macroeconomic indicators. They focus on one thing only: extracting accurate information from their designated "data warehouse" in the most efficient way possible. This "single responsibility" design significantly improves answer precision (signal-to-noise ratio).
3. Answer Aggregation: The results from multiple DRAs may include structured data (like numbers) and unstructured text (like report summaries). The aggregation layer within the router performs a "reduce" operation, intelligently combining these disparate answers into a single, coherent, complete, and source-cited final response for the user.

LangGraph is the linchpin here. It provides the tools to build this complex workflow (like state management and conditional routing) and allows the development team to easily iterate and test the entire routing logic locally, greatly boosting development efficiency.

Trend Insight: From "Monolithic Models" to "Specialized Collaborative Systems"

Kensho's practice reveals a deeper trend in AI application development: Future complex AI systems are unlikely to be a single, all-knowing "super-brain." Instead, they will likely be federations of multiple specialized "sub-brains" working in concert. This mirrors how human organizations function—we have doctors, lawyers, and engineers collaborating to solve complex problems, rather than expecting one person to master every field.

This trend implies:

• Shift in Engineering Complexity: The challenge partially moves from "how to train a better model" to "how to design efficient agent collaboration protocols and architectures."
• Higher Walls for Data & Domain Expertise: Owning high-quality, structured proprietary data (like S&P Global's) combined with an effective AI architecture creates a deeper moat.
• "Trustworthiness" as a Core Value: In high-stakes fields like finance, healthcare, and law, the traceability of AI outputs and the authority of data sources are often more critical than raw "intelligence." Grounding's citation feature is a direct response to this.

Practical Value: Lessons for Developers and Enterprises

For developers and enterprises building AI applications, Kensho's case offers several takeaways:

• Architectural Choice: In complex, multi-source data environments, consider a "route-dispatch-aggregate" multi-agent pattern over trying to solve everything with a single LLM call.
• Protocols First: When working across teams or integrating multiple systems, define clear data exchange protocols (like Kensho's DRA protocol) early to avoid significant integration pain later.
• Focus on "Trust-by-Design": If your application operates in a professional domain, make "data provenance" and "answer citation" first-class citizens in your system design from the outset. This can dramatically increase your product's credibility and commercial value.
• Leverage Existing Toolchains: Frameworks like LangGraph provide a solid foundation for building complex agent workflows, saving you from reinventing the wheel.

The Counterintuitive Insight

One point that might be overlooked is that the core goal of this system is not to make the AI "smarter" but to make its outputs "more trustworthy" and "traceable." In finance, a seemingly clever answer without a source is worthless or even dangerous. Kensho invested heavily in protocols and architecture primarily to ensure that every number and every conclusion can be traced back to an authoritative S&P Global data source. This reminds us that in enterprise applications, engineering rigor (protocols, architecture, audit trails) is often as crucial as the intelligence of the model itself.