Several rounds of ordinary Agent-mode prompting have produced plausible patches for one bug, each changing behaviour without resolving it. What does Cursor identify as the difference Debug Mode brings to this situation?

The principle — What separates Debug Mode from ordinary Agent interaction is method, not horsepower. Cursor frames it as a different approach: when standard Agent interactions struggle with a bug, Debug Mode uses runtime evidence rather than guessing at fixes.

Why the key is correct — A bug whose cause cannot be read out of the source will not yield to more reading of the source, however the request is worded. Debug Mode changes the input instead of the effort. It forms hypotheses, instruments the code with log statements, has you reproduce the failure, and reads what actually happened at runtime before making a focused fix. The patches stop being plausible speculation because they are chosen against observed behaviour, which is why the mode is recommended for exactly the cases that resist ordinary prompting: race conditions, timing issues, performance problems and regressions where something used to work.

Why the others are wrong — Reaching for more context is the capability assumption wearing a different hat; a wider view of the same static text still contains nothing about execution. Regenerating patches and keeping whichever moves behaviour the most is trial and error made faster, and it reliably produces changes that hide a symptom while the cause survives. And consulting issue trackers or past conversations is not what the mode does, since its evidence comes from your running program rather than from reports of similar-looking failures.

Remember this — When speculative patches keep missing, change the evidence rather than the prompt.

Sources — Cursor's Debug Mode documentation.