Grep beats LSP? Why coding agents ignore your fancier tools
Summary
A study compares lexical search (grep) with LSP-backed semantic navigation for coding agents, showing that tool harness design and interface shape significantly influence model behavior and task success. The findings indicate that semantic retrieval can offer precision gains in noisy code, but effectiveness depends on the task, codebase, and how results are exposed to the model. The piece also discusses ACP and AgentConnect as approaches to maintaining native runtimes and meaningful tool surfaces.