They work best when the codebase is reasonably organized and documented. A clear structure, a readme, and existing tests give the agent the context to make good changes and to verify them. Messy, undocumented code makes an agent guess more, which raises the error rate.
For big or unusual codebases, agents still benefit from guidance: point them at the relevant files, describe the conventions, and review changes closely at first until you trust how the tool behaves in your project.
Updated July 2026