Chaotic, unreliable results from tools

[bulyaki (Peter Bulyaki)]

Where does the bug appear (feature/product)?

Background Agent (GitHub, Slack, Web, Linear)

Describe the Bug

I am experiencing a sharp decline in the accuracy of agentic responses, something is not normal, broken, not working properly. I am on MacOS, with latest cursor release. I am in auto mode, but I make the agent announce the model and version every time. The issue is not dependent on model type ot version at all, I get the same issues from Claude 4.5 too. Examples:

• Tool tells me a location for a log file it created, it also gives me a command I can copy/paste to monitor the logs (something like “tail -f …”). But the location is wrong. E.g. it gives me ~/Library/Application Support/Something/Somethingelse, then it creates a log in ~/Library/Something/Somethingelse. And after this when I run the code and return, it can’t find its own log file, and starts to investigate the code why the log file was not created - but it does not notice the incorrect path.
• It gives me instructions for something at the finishing summary of the agent’s work, and puts a completely irrelevant piece of code there instead of the one it is talking about. E.g. it is talking about checking the logs in the log file, and you expect a piece of code for that, but it repeats the same code it has already shown you, that resets the cache.

These are major issues in my opinion, it makes you lose trust even in the smallest bit the agent is saying.

Steps to Reproduce

There is no exact instruction for this. I can share entire conversations with you in private if necessary.

Expected Behavior

As above I mentioned, there are trivial bugs in agent responses, like telling the user a file path, but using a different one, giving the user a shell command that is irrelevant, etc…

Screenshots / Screen Recordings

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 2.0.43 (Universal)
VSCode Version: 1.99.3
Commit: 8e4da76ad196925accaa169efcae28c45454cce0
Date: 2025-10-30T18:49:27.589Z
Electron: 34.5.8
Chromium: 132.0.6834.210
Node.js: 20.19.1
V8: 13.2.152.41-electron.0
OS: Darwin arm64 25.0.0

For AI issues: which model did you use?

In auto mode, but Sonnet 4.5 mostly.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

[condor (Condor (Tee))]

hi @bulyaki and thank you for the detailed bug report.

I do not see a degradation of requests, could you share a bit more so we can see what is happening and why it happens so we can reproduce it:

• Was that on a longer thread or on a thread with large context?
• Does it occur on new chats as well or only on existing chats?
• Are many rules, logs etc. attached?

Could you post a Request ID with privacy disabled so we can look into the details? https://docs.cursor.com/troubleshooting/request-reporting

[bulyaki (Peter Bulyaki)]

I can’t reproduce it, after restarting Cursor it started working. I suspect that the reason may be that the agent output failed to ask for a confirmation to apply changes (the “Apply All” button did not show up), so Cursor thought that the previous chat was not yet finished. But it looked like as if it did finish.
When I restarted cursor at this point, it restarted with new output in the agent window, with an “Apply All” button at the bottom. After clicking on this button the copy/paste started working.

[condor (Condor (Tee))]

Thanks for the update. Glad it is resolved.