Interrupting the agent mid-work automatically reverts all work done so far (this is a bad, BAD idea)

Bug Reports
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Please, for the love of God, do not roll back all changes made by the agent if I interrupt it with the stop button! I have already lost important work multiple times because of this behavior.

Sometimes we just want to interrupt the model when it starts doing something we do not want, or when we need to examine something ourselves before letting the agent continue. In those cases, the work done so far is good, either fully or partially, and it should not be discarded.

If I truly want to roll back all changes, I can simply click on my last prompt and hit the rollback icon. Done!

Let me explain why this behavior is such a bad idea:

Not only can I lose significant work that took many minutes to generate, but worse, sometimes I do not even realize that the work has been rolled back. After a few more prompts, and possibly a lot more new work done by the agent, I notice that the code no longer works correctly or that bugs previously fixed have reappeared. That is when I realize the earlier work was lost. But by then, I no longer know exactly which changes disappeared, and I waste a lot of time trying to identify and restore them.

It is extremely frustrating, and a considerable waste of both time and credits.

And please do not respond with “Just remember not to do that” (Steve Jobs style) or “Just re-run the last prompt.” It is not that simple, for all the reasons I have explained. It is easy to forget, and no other IDE or agent tool behaves this way, thankfully.

I understand it might have seemed like a good idea when it was implemented, but trust me, it is not.

Steps to Reproduce

Easy:

  1. Write a detailed prompt that requires a lot of work.
  2. Let the agent start implementing it.
  3. After about half an hour, when the agent is just about to finish, think, “Wait! I need to check something before it changes this part!” and hastily click the stop button.
  4. Boom! All the work is gone!

Expected Behavior

Just interrupt the agent, don’t revert all changes done since the last prompt.

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 2.0.60
VSCode Version: 1.99.3
Commit: c6d93c13f57509f77eb65783b28e75a857b74c00
Date: 2025-11-05T00:50:54.645Z
Electron: 37.7.0
Chromium: 138.0.7204.251
Node.js: 22.20.0
V8: 13.8.258.32-electron.0
OS: Darwin arm64 24.6.0

For AI issues: which model did you use?

Sonnet 4.5

Does this stop you from using Cursor

No - Cursor works, but with this issue

2

Hey, thanks for the report. This is a real problem, and your explanation of why the current behavior is frustrating is completely clear.

For now, if you need to pause the agent to check something, try making small git commits to track progress. I know it’s not ideal, but it reduces risk.

The team will review everything. Thanks for the detailed description of the problem!

1 Like
4

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

When using the Cursor agent to make edits, it seems that the apply model keeps reverting the changes that the agent is making, requiring many more tokens, more cost, and more time to make these edits. This is not a one time thing, it has been happening for at least the last few weeks, quite frequently.

example chat attached

Steps to Reproduce

It seems to happen mostly when you’re doing a multi-step edit with a somewhat dumb agent, i.e., Grok code or Claude Haiku 4.5.

Expected Behavior

The agent should be able to make multiple different edits throughout a single agent message chain without it reverting the agent’s previous edits.

Screenshots / Screen Recordings

Screenshot 2025-11-05 at 2.53.38 PM.png
Screenshot 2025-11-05 at 2.53.38 PM.png758Ă—974 72.2 KB

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 2.0.60 (Universal)
VSCode Version: 1.99.3
Commit: c6d93c13f57509f77eb65783b28e75a857b74c00
Date: 2025-11-05T00:50:54.645Z
Electron: 37.7.0
Chromium: 138.0.7204.251
Node.js: 22.20.0
V8: 13.8.258.32-electron.0
OS: Darwin arm64 24.6.0

For AI issues: which model did you use?

Grok Code, Haiku 4.5

For AI issues: add Request ID with privacy disabled

2586600b-9957-4e1f-90e6-376f2347db2b

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

5

Hi, thanks for the kind response and for looking into it.
Meanwhile, I’m committing at every step, for fear of losing something.
I hope this gets solved soon.

1 Like
6

Thanks for the reply! So far it seems to work well to accept changes throughout a chat, however it often occurs during 1 agent response, ie I tell it to do something and while its trying to do that one thing it gets somewhat stuck in a loop.

Overall definitely not a deal killer and I love the continuous improvement that the team is making to Cursor!

1 Like