Which AI coding agent works best for large codebases?

No public documentation we reviewed ranks AI coding agents on large codebases, and results depend heavily on the repository, the task and the model. This thread-starter lists what to examine rather than naming a winner.

Large repositories exceed what any model can read at once, so how a tool manages context matters. Cursor documents Max Mode, which allows larger context windows, up to 1M tokens, on certain models. Claude Code documents a CLAUDE.md file that loads project instructions each session, and subagents that work in their own context windows and return a summary.

Instructions and conventions matter as much as raw context. Check how each tool lets a team share rules, commands and conventions so that changes follow your architecture.

Cost behaves differently at scale. Cursor bills included usage at model API rates with on-demand usage after it, and Claude Code subscriptions use rolling usage windows. Long sessions over large repositories can consume allowances quickly, so measure real usage.

Oversight also scales with change size. Claude Code documents permissions, sandboxing and checkpoints; the Cursor sources reviewed did not document agent permission controls, so verify them directly. Review effort for large diffs is often the real bottleneck.

The most reliable method is to run the same cross-module task with tests on a branch of your own repository in each tool, then compare correctness, review time, cost and developer preference. Reader replies are not open yet; see the Cursor vs Claude Code comparison for a side-by-side summary.