In my comment on the bug report, I explained that the system terminated the Subagent with an error indicating a model overload on the Cursor servers.
I find it strange that, in my case, the Subagent which had been running for more than 20 minutes was terminated, rather than another user attempting to make a new request.
I believe that when the load is too high, it’s better to prevent new users from submitting requests rather than interrupting sessions that are already in progress.