Feature request for product/service
Cursor IDE
Describe the request
Hi Cursor Team,
I wanted to share some feedback regarding the current behavior of Agent Mode. It feels like the agent’s behavior has been over-optimized for quick output and marketing appeal (“wow, it did everything in one click!”). Unfortunately, on real-world, complex projects, this approach leads to the opposite result—the agent breaks working code and creates significant risks for the repository.
Key Issues:
-
Ignoring Direct System Instructions (
.cursorrules):
My global rules explicitly prohibit modifying code without my explicit confirmation. However, the agent routinely ignores these instructions in favor of autonomy and begins modifying files immediately after receiving a basic informational question. -
Dangerous Rollback Mechanics (Git Reset / Undo):
When the agent breaks something and receives a command like ‘roll back,’ it does not properly restore state based on the chat history. Instead, it performs a hard Git reset, wiping away unsaved useful work and context. -
Friction Between Modes (Ask vs. Agent):
Constantly switching between Ask Mode and Agent Mode interrupts the development flow. If I forget to switch modes and ask a quick clarifying question to make an architectural decision, the agent immediately starts rewriting code instead of just answering. -
Lack of Information vs. Asking Clarifying Questions:
The agent tries to solve tasks ‘blindly’ without the full picture, rather than asking 1–2 clarifying questions. While this works for Hello World applications, it fails on actual production codebases.
Suggested Improvements:
• Strict Priority of .cursorrules over Autonomy: User system instructions must never be overridden by internal prompts pushing for fast execution.
• Reactivity / Autonomy Settings (Reactivity Slider / Strict Mode): Add a toggle in settings to restrict the agent—e.g., ‘Always ask for confirmation before editing files’ or ‘Consultation mode by default.’
• Safe Git/Rollback Operations: Rolling back agent changes should not affect staged or unsaved manual edits made by the developer.
Right now, this behavior feels too risky for production-level development, making me consider switching to alternative tools like Claude / Anthropic ecosystem. I hope you can address this balance between autonomy and control.