RESEARCH · RESEARCH · #1738
NVIDIA KGMON places second in KDD Cup 2026 Data Agents with constrained agent harness
The NVIDIA KGMON team placed second in the KDD Cup 2026 Data Agents competition by building a tightly constrained agent harness around a small, fixed LLM. Key techniques included unifying structured data into a single SQLite interface, a preflight schema-scouting step, a minimal toolset (schema(), sql(), write_answer(df), prose_helper()), middleware to repair malformed calls, separation of document inspection via zero-temperature LLM calls, and execution-trace logging to diagnose failures; recorded sessions and slides from the competition are in the KDD presentation archive.
KEY POINTS
- The NVIDIA KGMON team placed second in the KDD Cup 2026 Data Agents competition by building a tightly constrained agent harness around a small, fixed LLM.
- Key techniques included unifying structured data into a single SQLite interface, a preflight schema-scouting step, a minimal toolset (schema(), sql(), write_answer(df), prose_helper()), middleware to repair malformed calls, separation of document inspection via zero-temperature LLM calls, and execution-trace logging to diagnose failures; recorded sessions and slides from the competition are in the KDD presentation archive.
- This result highlights that reliability for data-analytics agents can come more from a smaller, well-structured harness and preprocessing than from larger or more open-ended models, a practical lesson for building systems around smaller open models.
WHY IT MATTERS
This result highlights that reliability for data-analytics agents can come more from a smaller, well-structured harness and preprocessing than from larger or more open-ended models, a practical lesson for building systems around smaller open models.