Has anyone had ANY luck with the auto mode? Especially in agent, it seems to hallucinate hard, forego any semblance of best practice, and often implements disastrous changes. I refine my prompts pretty hard for it because of this, but still the results are usually so bad I’d rather just copy+paste code into an AI elsewhere.
It certainly does not seem like these get routed to decent models. Is this user error or have others found Auto to be useless?
Yeah, it’s useless
It seems to be pretty mixed. You have a lot of people in the forums who hate it, but you’ll find a decent amount of people like me who love it.
it might just be user error but i’d have to know how your projects and prompts compare to mine.
And before people say their projects are more complex than mine, they probably aren’t.
They say practice makes perfect, but in reality useful good practice makes perfect.
Show us what you do then? Other wise it’s completely useless unless you sit there and monitor it 95% of the time to ensure you delete its chopped up coding. Also when you ask what model it uses what response does it give? Please provide images.
Gemini 2.5 Pro will frequently refuse to tell you what model it is. But it does use Claude 4 Sonnet sometimes. I only have 2 instances open right now, on more basic projects, but it is using Claude 4 Sonnet on one of them.
It isn’t even close to being done yet, but I have it periodically working on a video game I have always wanted to play. I also have it working on programmatic video generation using Ollama’s llama3, I got image generation working programmatically AKA not using stable diffusion but using blender instead.
Auto is not perfect it makes a ton of mistakes, but the secret is to have enough instances open, that it doesn’t matter how long it takes.
Open 50 instances of cursor and then ask it on auto in a new chat every time, which model it is using. It does use Sonnet 4 sometimes.
Auto is of no use, i can use chatgpt 3.5 and it would give better results than auto 
sometimes it works very well honestly.
Auto is absolutely useless, except for a very simple task on single file and not too long. Rather than using auto I prefer to copy parts of code and ask chatgpt. But of course it is very uncomfortable.
I’m running out of the pro license I signed up for, usually in 2.5 weeks. That makes my pro license unusable half the time.
I’m very unhappy about this situation.
I’m out. Not paying anymore for this.
Auto is bad, only useful for easy tasks. Feels like betrayel to stuck with auto
Why do you think people here seem to struggle so much with auto?
I think it has a lot to do with the complexity of your prompts. I expected Sonnet-4 performance from a prompt while in Auto-Mode, and what I got was gibberish. The sonnet-4 performance allowed me to provide a prompt, such as ‘produce a plan to implement RBAC,’ and it enabled me to gather the 11 comprehensive steps needed to implement it and complete each one.
Auto Mode seems to work if you’re doing little steps at a time.
For me, I cancelled my Cursor subscription because I realized that as soon as my tokens to access Sonnet-4 prompts had run out, the subscription prevented me from continuing to program for the month. That’s no way to work. I’m currently on the bleeding edge with Code; laughs. But as a senior dev, I love it.
Really? I give auto a prompt of “Continue working” and it has made pretty complex things based on nothing more than the name of the project file. I get even better results if I put in the bare minimum work of listing all my requested features and requirements.
I also cancelled my sub. It’s too easy now to blow through credits and using auto isn’t possible.
Starting the last few days, it acts really bad and the code generator is pretty slow. I went to shave my beard and cursor was not able to finish generating an index.html page. Let`s not talk about hallucinations. And i ask it to change one part of the code and he is starting to change the whole structure of the page… Acting like an real idiot, and i heard people and companies that fired their human workers in favor of this retarded AI. I have 5 days left until renewal and i will stop paying for such of piece of garbage AI. I have not found yet an AI to be able to help me with websites design, without messing around with it. I tried Github Copilot also… Worst than Cursor at generation, but way better on speed…
Wow Nice Brother
Yes. Auto is not very good at anything unfortunately. Claude 4 Sonnet Thinking is my choice and yes it costs extra.
Auto is completely junk! Here is a comparison of what auto thinks is useful for a test and what Claude thought of that test. They should really replace auto with a better alternative, it is worse than useless because it’s a waste of everyone’s time.
You asked two different questions and got two different answers. Why are you surprised?
you missed the forest for the trees. Auto wrote the test after I gave it a complete guide on writing tests in context and I knew it was ■■■■ and it didn’t follow instructions; I just wanted to see if it had any idea of what it produced. Having used auto many times it is just inferior to sonnet in every way. If you have auto write tests it will find every way to make the test work (true==true) without actually creating a useful test to the point of not actually using the features your testing and instead testing the underlying libraries. If you’re happy with auto, use it and be happy 
. Meanwhile I’ll use a model that has a clue about software.
