Repo maps let agents navigate large codebases

A repo-map summarizes codebase structure so the model navigates large repos without full inclusion. The agent sees what exists and where, then reads only what it needs. This keeps the context (see: context management is the dominant variable in agent quality and cost) window focused on the actual change.

Without a repo-map, the agent’s default strategy is to search broadly, read many files, and fill the context window with exploration before reaching the actual task. This burns tokens and degrades quality as the window fills. The repo-map inverts the cost curve: a tiny structural summary up front saves thousands of tokens of exploratory reads. The pattern generalizes to any tool that provides a compressed view of the codebase; the principle is always the same: give the agent the map, not the territory.