NVIDIA Team Takes Second at KDD Cup 2026 Data Agents: Lessons in Reliable Agent Design

It is a technical challenge that asked agents to answer natural-language questions over heterogeneous data sources including databases, CSV and JSON files, prose documents, PDFs, and briefing videos.

The team focused on making the agent's harness smaller, clearer, and easier to verify rather than adding complexity, and placed second.

The approach is immediately usable as a design guideline: start with a simple, verifiable agent harness before adding complexity, especially in the heterogeneous data environments common across the region's government and enterprise sectors.

Conventional agents tend to grow complex pipelines with many tools, while the KGMON approach shrinks the harness and prioritizes clarity and verifiability, making errors easier to trace and reliability easier to raise.