A new arXiv paper tests whether language servers actually make AI coding agents more token-efficient than plain lexical search. The authors argue that the claim is often repeated but rarely measured under equal task-success conditions.
The study compares semantic retrieval through the Language Server Protocol with grep-style retrieval using a tokens-to-success metric across Python and TypeScript repositories. Its preliminary result is conditional and often negative: for symbol-named localization, language-server access increased token use by 6% to 118%, and agents frequently ignored it when it was available. For reference-completeness tasks, semantic retrieval improved precision but did not reliably save tokens or overcome limits in agent thoroughness.
The finding does not mean language servers are useless for agents. It means their value depends on the task, setup cost, and failure mode. For teams building coding-agent tools, the paper is a reminder to measure full workflow cost instead of assuming that more structured retrieval automatically makes an agent cheaper or better.