the Gemini models are now doing this very often. I left them to fix a bug, but then they stay stuck in this loop repeatedly.
Hey Luiz_Gabriel_dos_San , It’s actually not only you. It’s also not an issue of the model or antigravity.
Gemini 3.8 Flash
The reason is that model always tries to get an accurate idea of all the files. So it reads all the files and every part of the code. It’s really frustrating because it exhausts the quota quickly over time. So it’s not only you.
You can check my post on this same problem,
https://discuss.ai.google.dev/t/gemini-3-6-flash-is-better-than-3-7-and-3-8/183802?u=sunerax
Gemini 3.7 Flash
From my experience 3.7 is the worst model in the gemini series. So don’t use it unless you want to waste quota or get unreliable results.
Gemini 3.6 Flash
I prefer the 3.6 model because it is really capable of making updates quickly with a few file readings. If it’s the initial prompt, it would read and explore the project a little, but after that it quickly make changes. Also it’s really good at doing things if we explain what to do correctly.
3.8 is also good, but eats quota quickly over time with file readings. So I actually work with 3.6 on my projects if I couldn’t find a bug or do a quick update. It’s not reliable as much as 3.8, but still saves quota. It’s better than eating quota for a simple thing.
Hope you would get an idea! Have a nice day!
Yeah, this is a more recent thing.
3.7 was actually perfect. It never got stuck like that. Even after the release of 3.8 for me. 3.8 from the get-go always had this problem. And it is not just “Trying to be accurate”; it actually gets stuck on a loop, visibly reading the exact same lines of the file over and over again unless you stop it. Now, in my example, it was the 3.7, which Is why I feel like is a new regression.
But this is different from just the model doing its due diligence and being a bit wasteful, and actually a bug/maladaptation of the training or harness that is failing to make it remember it already got that information, because it gets truly stuck. It really won’t stop. You have to stop it manually. You can visibly see it reading the same portion of the same file over and over again.
like this is the moment i stoped it, but it was running for 7 minutes already and accomplished nothing, but kept reading the same 30 lines of the same file non stop. This was just this has been happening a lot. that’s just the point where i decided to put it out in the open.
Oh! I had the same issue with 3.8 and 3.7 but I let it continue for a while and after it finished I checked, the reason isn’t clear since the 3.8’s performance is at another level for coding, but it really did a great job.
It’s not only you. I had the same issue. Well, letting it continue was good for me but uses quota more. But in your screenshots, it seems like there’s a loop!
It’s hard to say whether it’s a bug in antigravity or model.
Could you please share few details to reproduce this issue:
-
The antigravity version you are using?
-
The operating system you are using?
Antigravity was on the 2.15.0 version.
Windows on the latest version possible for that day. Windows 11.
Could you please update to the latest version Antigravity 2.15.1, and let us know is the issue resolves or the issue still persists?
Oh, okay, maybe I wasn’t very clear about the timeline.
This has been a problem ever since Gemini 3.8 Flash came out. So, let’s assume I was on 2.12.2 (I’m always up to date with Antigravity). The problems started there; however, Gemini 3.7 was still fine. I lost some quota progress, but then I reverted to using Gemini 3.7 Flash, which until now had not done this or had this problem for me.
However, very recently, with the 2.15.0 update (or at least at that timing, maybe not because of the update itself but possibly some Google-sided rearrangement of resources in their cloud infrastructure), Gemini 3.7 Flash started to display those problems as well, on top of the 3.8 Flash, which was showing that problem from the outset.
That incident was what motivated me to post it here.
Now, the latest screenshots from the 45m incident were already in the 2.15.1 version. I tried to use the 3.8 again, but it still fails me catastrophically, even for simple tasks. (The prompt is on the screenshots.)
