I triggered a Cursor agent yesterday to convert a collection of scripts into a macOS desktop application. That was over 14 hours ago… and it’s literally still running.
I’m watching token consumption steadily increase, and no final deliverable has been produced.
I’m not even being billed yet (still within Pro usage), but this is obviously unsustainable. I’m concerned that if I walk away for a few days, I’ll come back to a $50+ token bill for a process that never completed.
So here’s my situation:
-
I asked Cursor to “turn my scripts into a macOS app” (a modest scope, I thought).
-
It has been running overnight and beyond, executing retries, video rendering attempts, and token-consuming calculations.
-
Cursor CLI still shows it running.
-
I’ve received no success message, and no new artifacts or outputs have appeared.
Questions:
-
Should I just kill the agent?
- I feel like I’m wasting tokens by the second, but part of me wonders if it might “finish” soon…?
-
Has anyone else experienced this?
- Where an agent goes into a “zombie” state, endlessly consuming cache reads and re-processing steps?
-
Any best practices to avoid this kind of runaway token spend in the future?
- (Model selection? Session limits? Agent safeguards?)
Suggestions for the Cursor Team:
-
Option to auto-terminate long-running background agents after X hours or Y tokens.
-
At minimum, a live token counter or alert when things are spiraling.
-
A dry-run estimate: “This task might consume X tokens” before launching complex agent jobs.
TL;DR
Cursor has been burning tokens for 14+ hours straight on an agent task that hasn’t delivered anything. It’s still running right now. Should I kill it? Has this happened to others?
