…which is still not provided 
Actually, losing control is one of the main reasons why I haven’t touched the background agents yet. Considering how closely I need to monitor the Agent in the Chat, I’m afraid to imagine what the ■■■■ the agent will do without supervision.
Don’t you think that’s a problem cursor should look into?
Considering how broken Grok is and what problems Gemini has with edit_tool, I’m ready to assume a failure on the Cursor side. But it is also necessary to take into account the level of current LLM. And the fact that you can change one word or its form in the prompt and your whole answer will change.
But again, I’ve never even run Opus, not even once, even with Pro+, and I was practically groaning when testing o3-Pro. And I rarely use Claude either, because I know it’s expensive.
Agreed on all, this why somekind of safeguards should be in place. The models are not there yet, they should be made available accordingly. If they have propensity to run away because of this or that, we should have some kind of alert or something. I’ve been using cursor for sometime now and I really want them to succeed but if people get penalized by loosing all their tokens because the request was not spot on it’s not the best image.
The second Opus that will constantly check the git diff of the first one, huh?
kimi k2 can’t come soon enough I guess
Yeah, I’m pretty disappointed with the reality of working with Cursor lately, from a cost perspective (well, and performance, honestly). I’m going to stick it out for now, but I definitely have my eyes and ears open for alternatives again. I don’t know if anybody else has found this, but its code seems to be growing increasingly erratic; it needs a ton of babysitting. I realize this may be the LLMs themselves, but still… At least I might pay less for the madness.
Same experience here. I’ve had very small instructions completely fail and random generations that were specifically requested NOT to happen. I’m not trying to vibe code a project, im trying to import a data model. This isnt rocket science and the “Auto mode” is failing time and time and time again.
mmmmm. its kind of a ■■■■ shoot- CC for heavy lifting- Roo is better than cline.
Hey, Kimi is already available in Cursor.
You’re not wrong. This is a product issue. Even Claude Code has a popup every $5 you spend in a session it pauses to warn the user.
Cursor has no such guardrails.
When a car manufacturor builds a car with faulty brake lines, they cannot blame the customer for driving too fast that they weren’t able to gently roll to a stop when the brake line exploded.
You can have cursor run in ask every time mode. You can have a simple python script keep an eye on usage. As far as i can tell this is user error. He even picked the most expensive model, in the mode that is designed to be used without looking at it. If you want specific results you shouldn’t be using back ground agents.
You know you can set agent to wait for you to approve every subsequent action ?
Letting it run of untethered on its own is dangerous, do that in auto mode if you have to.
Free run agents are often reckless even in the best case scenario
Review your cursor settings here:
Just try claude code any you will ask cursor for refund the next day
What behaviour do you want? Do you want it to abort the opus requests when it hits $10?
I’m annoyed by the changes they’ve made, but what they are doing is hard. There’s just not a lot of expertise in how to manage what the LLM does. Perhaps like MCP, they need a way to signal how much resources the model should use. I haven’t used opus, but I use sonnet all the time. Interestingly it looks like you got limited very quickly, but I’ve slowly been using my credit, using auto and then switching to claude for bigger more important changes, and I’ve gone over the level where you have stopped.
If I had to guess, because of the speed at which you’ve “burned” through your credit, I think they’ve stopped your usage earlier than for other users. I think this is because they are currently (in my mind, incorrectly) trying to limit users who burn up a lot of credit quickly. I’m guessing some of those users would cycle accounts if there was a set level of credit, and so they are limiting people who hit it more quickly, than those who take some time and built it up slowly.
I think this is also why they are being a bit unclear on exactly what you are entitled to. Because if they are clear on the algorithm, then the same people who are trying to maximise the value they squeeze out of their $20, will run 10 accounts and abuse the mechanics.
But for now, based on me and my team’s usage, if you have a slow and steady approach, you seem to get more credit. Also if you use cheaper models. I’ve only tried opus once, and heavily restricted what it could do, and I watched it executing.
Loving GitHub Copilot. It’s bug-free, affordable, and doesn’t play games with usage. No complaints.
Agreed. I’m currently using both Cursor and VSCode with GitHub Copilot in parallel ($10 for unlimited GPT-4.1 and 300 Sonnet requests).
GitHub Copilot codes quite fast and doesn’t generate much unnecessary chatter.
I still use Cursor when doing a lot of manual coding. Cursor Tabs are still excellent.
This is something I love with CC. The way it controls the context + forcing you to review everything step by step. I think just another mode for Cursor to build if they want to.
