Where does the bug appear (feature/product)?
Cursor IDE
Describe the Bug
When using Multitask to delegate small pieces of work to background agents, a subagent may need approval to run a shell command. After switching back and forth between the main thread and the subagent thread, the permission prompt can disappear without being approved, leaving the subagent stuck waiting indefinitely. At the same time, the side pane does not indicate that the subagent is blocked on a permission request, so there is no clear signal that the thread needs attention.
Actual behavior:
The Allow button / permission prompt may disappear even though it was never clicked.
The subagent remains blocked waiting for permission and cannot continue.
The side pane agent list does not show any badge, warning, or other indication that this subagent requires attention.
Steps to Reproduce
Start a background subagent using Multitask.
Let the subagent attempt a shell command that requires user approval (Allow prompt).
Navigate between the main thread and the subagent thread a few times.
Return to the subagent.
Expected Behavior
The permission prompt should remain visible until it is explicitly approved or denied.
If the prompt is dismissed or hidden for any reason, it should be recoverable when returning to the subagent.
The side pane should clearly indicate that a subagent is waiting for permission or otherwise requires user attention.
Operating System
MacOS
Version Information
Version: 3.5.33 (Universal)
VSCode Version: 1.105.1
Commit: aac81804b986d739acab348ed96b8bea6e83cc50
Date: 2026-05-22T06:47:48.039Z
Layout: glass
Build Type: Stable
Release Track: Default
Electron: 39.8.1
Chromium: 142.0.7444.265
Node.js: 22.22.1
V8: 14.2.231.22-electron.0
OS: Darwin arm64 24.6.0
Does this stop you from using Cursor
Sometimes - I can sometimes use Cursor
This is a known bug. The permission prompt can lose its rendering state when you navigate between the main thread and subagent threads in Multitask mode, leaving the subagent stuck with no way to approve the command. Our team is aware of this and actively working on improvements to the approval flow.
For now, a couple of workarounds:
-
Enable auto-run – If your workflow allows it, switching to a more permissive auto-run mode (Cursor Settings > Agents > Auto-run) will skip the approval step entirely for commands that match your allowlist or sandbox settings. This avoids the stuck state altogether.
-
Start a new chat – If a subagent gets stuck, starting a new chat session and re-running the task should work since the approval state resets.
Regarding the missing attention indicator in the side pane: you’re right that there’s no visual signal when a subagent is blocked on approval. This is a known gap and there’s active discussion around improving how subagent states are surfaced.
thanks.
It happened now again, but something worse happened.
I clicked on the back arrow to minimize the sub worker dialog, and the allow button disappeared. I opened the sub dialog again just to understand at which point it got stuck and I saw that the previous command that require permissions started running and finished. I’m sure I did not click on the allow.
i replied a few days ago from my email, as the email suggested i can do, and now that i visited the site, it did not have the reply from my email…
so i posted it above…Thought you’d like to take care of it as well 
Thanks for the additional detail. The behavior you’re describing - where a command ran and finished on its own after you minimized and re-opened the subagent dialog - is actually a different issue from the original prompt disappearing. The original bug causes the subagent to get stuck, what you’re describing is a command executing without your approval, which is a more serious concern.
Could you create a new thread for this specific issue? That way our team can track and investigate it properly rather than it getting buried here. In the new thread, please include:
-
A screen recording if it happens again - showing the subagent waiting for approval, you clicking the back arrow, and then the command having completed on its own
-
Your auto-run settings (Cursor Settings > Agents > Auto-run)
-
Your Cursor version (Help > About)