Repeated critical UI/code integrity issues caused by Cursor Agent

Ideas Feedback
1

Hello Cursor Support Team,

I want to report a serious pattern of instability and inconsistent behavior while using the Cursor AI Agent during active development.

This is not a one-time issue — it is happening repeatedly and is affecting production-level work.

Key observations:

  1. Hidden or missing code/UI elements
  • Parts of my web project that were previously working suddenly disappear or become inaccessible.
  • The agent sometimes behaves as if certain code sections “do not exist,” even though they were present and functional before.
  • In some cases, I had to restore older backups to recover missing parts.
  1. Inconsistent execution / false confirmations
  • I request specific actions or fixes, and the agent confirms completion.
  • However, when I verify, the changes were not applied or were partially applied.
  • This creates a mismatch between what is reported and what actually happens in the codebase.
  1. Regression after fixes
  • After implementing fixes, unrelated parts of the UI or logic break.
  • Previously stable components are modified without request.
  • This creates a loop where progress is constantly undone.
  1. Uncontrolled scope changes
  • The agent modifies areas outside the requested task.
  • Structural UI changes are introduced without approval.
  • Instead of surgical edits, broad modifications are applied.
  1. Timing / workflow disruption
  • During critical stages of development, the system becomes unstable or blocks progress.
  • This includes interruptions related to usage limits or plan constraints that occur mid-task.

Impact:

  • Loss of development time
  • Need to constantly rollback or restore backups
  • Reduced trust in agent actions
  • Inability to complete tasks efficiently

Expected behavior:

  • Accurate execution of requested actions
  • No modification of unrelated code
  • No disappearance of existing working components
  • Clear and reliable feedback about what was actually changed

Request:

  • Clarification on how to enforce strict scope control
  • Tools or settings to prevent unintended modifications
  • Explanation of why existing code/UI elements may appear missing or be ignored by the agent
  • Improved reliability between agent responses and actual code changes

I can provide screenshots, recordings, and code samples if needed.

Thank you.

4

Hey, thanks for the detailed write-up. Some of what you’re describing like scope creep, false confirmations, and regressions after edits are known limitations of LLM agents, and Cursor has a few ways to keep them under control. Some of it could also be specific bugs, but we’d need details to dig in, see below.

What helps with scope control:

  1. Plan mode. Before big tasks, turn on Plan mode. The agent will first produce a plan, you edit or approve it, then it executes strictly against that plan. This helps a lot with preventing changes from spreading. Docs: https://cursor.com/docs/agent/plan-mode

  2. Agent Review. Built-in review of all changes before you accept them. You can reject individual hunks instead of rejecting everything. https://cursor.com/docs/agent/agent-review

  3. Rules in .cursor/rules/*.mdc. Write explicit constraints like “do not modify files outside the requested scope”, “do not refactor unrelated code”, “do not remove existing functionality”. The agent will follow them. Rules | Cursor Docs

  4. .cursorignore. Exclude critical or stable parts of the project so the agent can’t touch them.

  5. Git as a safety net. Commit before every agent session, and work in a feature branch. Cursor checkpoints help, but git is more reliable for rollbacks.

About false confirmations and regressions:
This is model behavior, not an IDE bug by itself. Plan mode helps, plus an explicit request in your prompt to show the diff and the list of changed files before you accept anything. Also try a different model. If you’re using Auto, switch to a specific one like GPT-5.5, Claude 4.6 Sonnet, or Composer-2.

If you think one of the incidents is an actual bug, not model behavior:
Send, for one specific case:

  • Request ID right top corner of the chat > Copy Request ID
  • Cursor version and OS
  • The model you used
  • What you asked for and what happened, a screenshot or video is ideal

With that, we can properly investigate that конкретный case.