NVIDIA KGMON Takes 2nd Place at KDD Cup 2026 by Slimming Down Its Agent Harness

NVIDIA's KGMON team took second place at the KDD Cup 2026 Data Agents competition. Agents had to answer natural-language questions, and the answers were spread across six data formats: databases, CSV and JSON files, text documents, PDFs and briefing videos. NVIDIA sums up its recipe in one sentence: make the agent harness smaller, clearer and easier to verify.

What problem the participants were solving

KDD Cup is one of the oldest data analysis competitions and is held alongside the KDD conference. The Data Agents track tests what breaks most often in corporate practice. Data does not sit in one tidy table. Some figures live in SQL, some in exports, some in a PDF report, and some are spoken in a briefing recording.

For an agent, that means three tasks at once. It has to understand the question, decide which sources to consult, and assemble an answer from fragments in different formats. For example, a question about an underperforming region needs both numbers from a database and an explanation from a text document. An error at any step turns a confident answer into a polished falsehood.

Betting on a small harness

The harness is all the code around the language model: the call loop, the set of tools, prompts, rules for routing between sources, and result checking. In agent systems it grows fast. Every exception and every patch adds a layer, and a couple of months later nobody can say at which step the agent took a wrong turn.

KGMON went the opposite way. Judging by the announcement, the team deliberately trimmed the harness and made each of its elements understandable and verifiable. For analytical tasks, this approach makes sense: an answer based on data has a source, and that source can be checked. The shorter the chain from question to answer, the easier it is to find where it broke.

An important caveat: the published announcement includes no score table, no name of the base model, no number of tools and no description of the architecture. All of this is presumably covered in the full write-up on NVIDIA's blog. Claiming more than the source says would be a stretch.

What the facts say

• Result: second place at KDD Cup 2026 Data Agents.
• Team: NVIDIA KGMON.
• Input data: databases, CSV, JSON, text documents, PDFs, briefing videos.
• Task format: natural-language questions, with answers to be extracted from heterogeneous sources.
• Solution principle: an agent harness that is smaller, clearer and easier to verify.

What you can take into practice

A company best known for GPUs is now showing its strength in applied agents too. But for practitioners, the principle matters more than the leaderboard position. If you are building an analytics agent inside a company, the lesson from KGMON is this: simplify first, then add complexity.

In practice, that means a few things. Evaluate the agent separately on each source type: quality on SQL says nothing about quality on PDFs. Keep the toolset minimal and remove anything that does not move the metric. Log intermediate steps so that any answer can be traced back to the original file. And compare a complex setup against a simple baseline: if the gain is small, an extra layer of harness only creates future problems.

It is worth waiting for the full technical description of the solution. That will show what exactly KGMON's "small" harness consists of and how well its experience carries over beyond the competition.