NVIDIA team places second in KDD Cup data agents

NVIDIA's KGMON team placed second in the KDD Cup 2026 Data Agents competition. In a technical blog post, the team describes a system built on one idea: make the agent's harness smaller, clearer and easier to verify.

Agents in the competition answered natural-language questions over mixed data sources, including databases, CSV and JSON files, prose documents, PDFs and briefing videos. Organizers required every team to use a small, fixed LLM. That left the harness as the main thing teams could optimize.

The team worked from two principles. The first was to constrain the action space. KGMON loaded all CSV and JSON files into the task's existing SQLite database, which gave the agent a single SQL interface. A persistent Python environment then exposed only a few helpers: schema(), sql(query), write_answer(df) and prose_helper(). Middleware repaired malformed tool calls so that one bad call did not end an attempt.

The system also ran a read-only "schema scouting" step before the agent started reasoning. This step briefed the agent on tables, likely join keys, units, null patterns, look-alike fields and row-grain issues.

For documents, the harness blocked full-file reads and allowed only limited previews and regex searches. prose_helper sent relevant chunks to a separate LLM call with temperature 0 and reasoning disabled. It returned either short answers or extracted tables, which kept raw text out of the main agent's context.

The second principle was to make every attempt inspectable. Execution traces, repeated attempts and trajectory inspection helped the team find failures. Shorter valid runs also left turns free for extra attempts and for ensembling.

NVIDIA notes that this is not a general recipe. It says table extraction from documents suited the competition's tasks, and the available excerpt does not give scores or the margin to first place.