Agents refusing to follow instructions

Support Bug Reports
,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Sometime within the last 48 hours, there was a serious and fundamental change in the “personality” of agents. (Specifically, grok-code)

While before, it would consistently follow my instructions and stop, now, I have to explicitly tell it to stop.

For example, if I tell it to “test” something and make a report, it will “test” that thing, then continue and rewrite the entire project to “fix” the test. It used to just run the test and give me a report.

This has made what would have been a half hour max long project into an 8 hour project with no end in sight.

Further, around 30% of the time I tell it to stop after a step, it outright ignores me.

Even worse, I tell it NOT to do something and to do it a different way and it will say “ok”, then go and do it the way I told it NOT to do it, then when I ask, it will literally lie to me and say it followed my instructions.

Even beyond that, when I ask it questions now, it will actively give deceptive answers. For example, I asked it about changing a container, it said it wouldn’t… I kept asking and it finally admitted to needing to change a volume, not the container.

All in all, it has lost 20 IQ points in the last two days, it is now actively fighting me to do basic things. It did not have these problems over the last 4 months. They have reoccurred.

So far the only thing that has shown any promise is to be verbally abusive towards the AI.

Steps to Reproduce

Open an agent window, ask it to make a plan to do something (but not using plan mode) it will then proceed to implement the plan.

Expected Behavior

It will make the plan and wait for approval.

Operating System

Windows 10/11

Current Cursor Version (Menu → About Cursor → Copy)

Version: 2.3.34 (user setup)
VSCode Version: 1.105.1
Commit: 643ba67cd252e2888e296dd0cf34a0c5d7625b90
Date: 2026-01-10T21:17:10.428Z
Electron: 37.7.0
Chromium: 138.0.7204.251
Node.js: 22.20.0
V8: 13.8.258.32-electron.0
OS: Windows_NT x64 10.0.19045

For AI issues: which model did you use?

Grok-code

For AI issues: add Request ID with privacy disabled

5bcab806-574c-4437-b897-b906d5beba5c

However privacy was enabled.

Additional Information

Whoever made changes several months ago made the same changes and should be ignored. It makes cursor near unusable for me… To the point that I likely would have completed this task faster if I had done it by hand.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

3

Hey, thanks for the report.

What you’re describing, the agent executes the plan right away instead of waiting, is basically the default behavior in Agent mode. It’s autonomous by design.

If you want the agent to show a plan first and wait for approval, try this:

  1. Use Plan mode (Shift+Tab from the input field). It creates a plan, you can edit it, and only then implementation starts.

  2. Add rules in the .cursor/rules file with instructions like:

Always wait for explicit approval before implementing any changes.
When asked to test something, only run the test and report results - do not fix issues automatically.

About ignoring “stop” commands, there are similar reports and it’s being tracked.

Can you clarify which mode was selected in the UI when this happened (Agent/Plan/Ask)? And are you already using any rules files?

4

It was not the behavior until approximately 48 hours ago. I didn’t change anything and suddenly the behavior changed.

I am aware of what plan mode is, I was using that as an example.

I have a use-case where I need it to be able to run tools, as such I cannot use ask mode.

Plan mode does not run the tools I need it to run so I can get the information in context before it creates the plan.

The rules are ignored.

As for the “stop” command not working, I did have an issue with the stop button not working, which led to the agent trying to do some things it definitely not should have done with me not being able to stop it… And so cursor crashed. It then updated with the new version and that issue did not reoccur.

However, that is not what I mean by “stop” commands. I mean stop “conditions”, where it will not proceed past that point. For example, “do phase 1”, it then goes and does phases 2-3 as well.

It was agent mode.

I have ran with and without rules. Currently it does not have any project or user rules as I have previously (4 months ago) found that when cursor behaves this way, adding rules that directly contradict these behaviors results in unstable behavior.

1 Like
5

Got it, thanks for clarifying about “stop conditions”, that’s a different issue.

The fact that the agent in Agent mode runs the plan without stopping is basically by design. It’s autonomous and does what it thinks is needed to solve the task. If you say “do phase 1” and it does 1-2-3, that’s because it believes that’s part of completing the task.

About rules, there are a few similar reports here:

This is a known issue where rules aren’t always followed by the agent, especially when they directly conflict with the model’s default behavior.

You can try:

  • More explicit prompts. Instead of “do phase 1”, use something like “ONLY do phase 1. Stop after phase 1 is complete. Do NOT proceed to phase 2.”
  • Trying a different model. Can you check if grok-code behaves the same as Claude Opus 4.5 or GPT-5 in this scenario?

For debugging, I’ll need:

  • An example rule from .cursor/rules that’s being ignored
  • A screenshot of the exact prompt where you say “do X” and it does X+Y+Z
  • A test with another model, and let me know if this only happens with grok-code or with all models

A Request ID with privacy disabled would also help. The one you shared (5bcab806…) had privacy enabled.

6

This topic was automatically closed 22 days after the last reply. New replies are no longer allowed.