AI Agent activated itself when I was away from my computer at lunch. Problem is that it consumed 40% of my premium tokens. According to usage logs it happened at Jul 1 at 12:28 PM used claude-opus-4-7-thinking-xhigh and consumed 24.1m tokens. the premium model was indeed preselected as I was working on something, but the last usage was at 10:55 AM so 1.5 hours before the automatic triggering.
Another problem is that I dont even understand what it did. The agent chat has no time stamps so I can’t verify that.
Steps to Reproduce
No idea, to be honest…
Expected Behavior
The agent dont activate itself automatically.
Operating System
MacOS
Version Information
3.9.8
For AI issues: which model did you use?
claude-opus-4-7-thinking-xhigh
For AI issues: add Request ID with privacy disabled
Hey, thanks for the detailed report. Cursor doesn’t start agents on its own, but what you’re describing matches a known pattern we’re tracking. If an Agent session was left mid-turn, it can resume after your Mac goes to sleep and wakes up, or reconnects to the network, and then keep running without you.
Then a second factor kicks in. In Agent mode, claude-opus-4-7-thinking-xhigh rereads the full, growing chat context on every step, so a long autonomous tool-call loop can rack up tens of millions of tokens very quickly. That’s likely where your ~24M came from.
To confirm the scenario, a couple questions:
Did your Mac go to sleep while you were out for lunch?
Can you share that chat thread or a screenshot of the end of the session? That’ll help us link the case to the right spot.
I don’t have an exact ETA for a fix yet, but I’ll log this internally.
On refunds for used tokens, that’s handled by the team at [email protected]. Email them and they’ll look into usage credit for your account.
And on missing timestamps in agent chat, fair point. That’s a separate UX gap, and I’ll flag it.
Thanks for your message. Yes, my Mac definitely went to sleep while the task was in the middle of the execution. To be precise, the AI asked me a few clarification questions that I could not answer without my architect, so I left the questionnaire unanswered to come back to it later. At some point the Mac for sure went to sleep. This corresponds to your descriptions.
How do you treat this? As a defect? Because my concern is that this is really not a desired behaviour because without answering the AI questions the implementation can be critically wrong. And without me proving the next step it becomes an uncontrolled execution. Not to mention the expensive one that consumed 40% of my premium tokens. I basically have no tokens on day one after my sub was renewed for the new month…
Yes, this is exactly the scenario we track as incorrect behavior. If an agent has an unanswered question AskQuestion, it should not resume the session by itself after sleep or wake and finish the task on its own. This is even worse on a high-thinking model that rereads the full growing context at every step. That’s how you end up with tens of millions of tokens. This is on our side, and we’ve logged the issue internally.
Your reasoning is correct. Without answers to clarifying questions, the implementation can go in the wrong direction, and that’s uncontrolled and expensive execution. I agree this shouldn’t happen.
A couple practical tips until there’s a fix:
If you leave a session with an open question and step away, it’s best to stop the agent Stop before your machine goes to sleep.
For unfamiliar or large tasks using opus in Max mode, it helps not to keep Auto-Run on for everything.
For token refunds, that’s handled by the team at [email protected]. Email them and they’ll check the usage credit for your account. One more detail, the Request ID you shared is for a different small request. The real run that caused this is d1e88498-a47c-4238-b45a-4266c8696c5f, include that in your email and it should be faster.
Cursor rejected refunding the tokens even though the defect is on Cursors side. Here is the response.
Initial response:
I’m really sorry, but we’re unable to refund any on-demand usage charges, whether the usage was intentional or not. These are valid charges, and we incur real costs from our model providers. This policy applies consistently to all of our customers.
I hope you can understand our position.
Best,
Sam
Response after I requested clarifications:
The API costs from model providers are incurred at the moment tokens are consumed, regardless of what caused that consumption. Since these are real costs already paid on our end, the policy applies consistently to all customers whether usage was intentional or triggered unexpectedly.
I’m sorry I’m not able to process a refund here.
Best,
Sam
That means businesses need to evaluate the risk of uncontrolled AI execution in certain scenarios and that Cursor would NOT take responsibility for the faulty product.
I get the frustration. This is a really unpleasant situation, especially when tokens were burned through no fault of your own and right at the start of a new cycle.
We’re treating this behavior as a bug on our side. The agent shouldn’t resume a session on its own after sleep or wake if there’s still an open AskQuestion, and it definitely shouldn’t continue the task autonomously on a high-thinking model. We’re tracking the issue internally, but we don’t have an exact ETA for a fix yet.