Build an Agent That Thinks Like a Data Scientist: How We Hit #1 on DABStep with Reusable Tool Generation

NVIDIA's Kaggle Grandmasters LLM Agent Research Team presents the KGMON NeMo Agent Toolkit Data Explorer, an autonomous data analysis agent that achieved #1 on the DABStep benchmark with a 30x speedup over Claude Code baseline. The architecture uses a three-phase approach: a Learning phase where a heavyweight model generates reusable helper functions from representative tasks, a fast Inference phase using a smaller model (Haiku 4.5) that leverages the pre-built library, and an Offline Reflection phase for quality control via reflection and group-consistency checks. On hard tasks (84% of the benchmark), the system scored 89.95 vs. 66.93 for Claude Code with Opus 4.5, while completing tasks in 20 seconds vs. 10 minutes. The key insight is separating upfront knowledge building from rapid inference, allowing smaller models to outperform larger ones on complex multi-step tabular reasoning.