Agentic Repository Mining: LLM Agents That Dynamically Explore Repos Match Pre-Engineered Context Classifiers

arXiv·low signal

An empirical study across four software repository classification tasks shows LLM agents that dynamically explore repos via bash commands match the quality of simpler LLMs given pre-engineered context. Across eight configurations, agent-based exploration compensates for missing context but at higher cost. Practical insight: for repo mining at scale, the engineering effort shifts from context curation to agent orchestration.