I recently started using Cursor after using Claude Code for a long time. Right away, I ran into a major issue with response quality.
By the second I start new session, the agent’s response became unusable [e.g., hallucinating code / ignoring instructions]. The context window was still very small, so context bloat was definitely not the issue.
When I directly asked the agent why it gave such a poor response, it literally replied:
“Because I optimized for answering fast instead of answering right.”
Has anyone else experienced the agent dropping reasoning quality for speed so early in a session? Is there a setting or model preference to enforce correctness over response speed?