Can data-analysis agents teach themselves new skills without human labels?

Training AI agents to perform data analysis typically requires expensive labeled examples. DataCOPE flips this: it lets agents explore autonomously, then learns which exploration paths were successful by checking if reports covered key topics or if different solution paths reached the same answer. The system uses these signals to extract and distill reusable skills—procedural knowledge that works across similar tasks—without ever seeing ground truth labels. On both report generation and reasoning-based analysis, this improves agent performance substantially.