Tag
CodeNib is a multi-view data system that serves repository context to coding agents by building reusable lexical, dense, and structural views per commit, enabling faster updates and efficient context serving.
Introduces Agent Retrieval Bench, a file-level benchmark evaluating how well coding agents retrieve relevant repository files during the context-acquisition stage. The benchmark includes 427 samples across 25 repositories and evaluates various retrieval methods, finding no single family dominates.
This paper evaluates whether repository-level context files like AGENTS.md or CLAUDE.md improve coding agent performance, finding that LLM-generated context files offer little benefit and may reduce efficiency, while developer-written files are better but still not clearly advantageous.