NVIDIA took second at KDD Cup 2026 by shrinking its agent's harness to four functions

The model was fixed by the rules. KDD required every team to run the same small LLM, so the harness was the main thing KGMON could optimize ^[4]. The tasks also asked for more than retrieval: the agent had to inspect the data, pick tools, reason across sources, write an answer file and avoid the traps that show up in analytical work ^[3].

All structured data sits behind one interface. KGMON converted every CSV and JSON file into a table in the SQLite database the task already had ^[5]. The agent reaches it through two functions, schema() and sql(query). The team built both into its own Python environment; neither comes from Python or SQLite ^[6]. Before the reasoning loop starts, a scouting pass reads the tables for possible join keys, duplicate names, look-alike fields, units, null patterns and row-grain problems ^[8]. The agent gets that briefing at the start of each task and skips a discovery turn ^[9]. According to NVIDIA, the single interface cut routing failures and wasted turns ^[7]. The scouting cut errors from wrong columns, missed joins and confusion about what an answer row should represent ^[17].

The rest of the action space is small. Besides the two query functions, the agent has prose_helper(), which answers from prose or extracts it into SQL, and write_answer(df), which writes the final answer atomically ^[10]. Every answer goes down that one output path ^[13]. An atomic write means the scorer sees a complete file or no file. It is the least glamorous function in the list and the one I would copy first. Middleware repairs malformed tool calls, so one bad call does not end an attempt ^[11]. The Python environment keeps variables between calls, so the agent can reuse intermediate results ^[12].

The budget all of this protects is turns. NVIDIA says short, valid attempts left more turns for additional runs and evaluation ^[16]. The team found its failures through execution traces, repeated attempts and trajectory inspection ^[15]. A small model that spends one turn finding the schema, one on a syntax error and one on a broken file write has that many fewer turns for analysis.

Treat the result as a claim about one workload. NVIDIA wrote that the playbook "isn't a general recipe for every data science agent" ^[14]. The harness sections of the write-up do not name the model, report a score, or measure each technique on its own, so the share of the second-place finish owed to any single change is unknown. None of the four listed functions is described as handling video, though briefing videos were among the competition's sources ^[20].

For the pattern to transfer, a few things have to be true in your setup. Your structured data has to load into one queryable store before the agent starts. Your model has to be fixed, or expensive to change. Your tasks have to be narrow enough that a handful of functions covers them. NVIDIA's suggested stand-ins for the temporary SQLite database are a virtualized query layer or a governed warehouse interface ^[18].

I think this is the right tradeoff when the model is not yours to choose. The post's own summary: "Reliability often comes less from making the model more open-ended and more from building the right harness around it." ^[19]