Grep beats LSP? Why coding agents ignore your fancier tools

A study comparing grep lexical search with LSP-backed semantic navigation tools found that coding agents overwhelmingly preferred grep despite its lower precision, choosing it 94-100% of the time on simple code-location tasks. The research reveals that tool effectiveness for AI agents depends not just on result accuracy but on "LLM-friendliness"—how well a tool provides sufficient context, uses familiar interfaces, and presents output in a format the model can readily use. The findings highlight that an AI model's capability is shaped by the entire tool harness available to it, suggesting that engineering tool integration for agents requires considering factors beyond technical precision.