I am a Pro subscriber. I encountered an infinite loop bug while using the Gemini 3 Flash model with the ‘Plan’ feature. Why is this happening, and will it consume my quota?
3 Likes
Same issue here. Gemini 3 Pro is really unstable these days. It can’t correctly analyze situations. Claude almost becomes the only option.
2 Likes
Checked and it’s not open.
We appreciate the feedback. It sounds like the agent got stuck in a ‘tool execution loop,’ which can happen when Gemini 3 encounters an unexpected environment error and tries too hard to fix it autonomously.
If you encounter this again, we recommend:
-
Stop the agent immediately to save tokens.
-
Often, loops are caused by ambiguous success criteria. Try restating the request with a stricter constraint, such as: ‘Attempt this fix once. If it fails, stop and report the error logs to me directly.’
1 Like
I encountered the same issue today. I checked the request this time. My request is very clear:
Ask agent to investigate and report three bugs. (I even asked it not to implement any fixes.)
That’s all.
Before the loop, I didn’t encounter any issues in the same dialog.
I don’t think it is a pure “ambiguous success criteria“ issue.
1 Like
same here, started yesterday, thought this might be because chat got too long, so started a new one. but no, flash went into repeated mad loops every few queries, then i switched to 3 pro high, also went into mad loop when implementing its own plan for a simple web page redesign, then switched to pro low to follow written instructions to build and deploy a flutter client, went nuts too. this never happened even few days ago, while still using antigravity 8 hours a day. no idea what triggered it, some code/config change on your side? a hardware failure on one “cluster” or whatever it is called? something else?
Mine had an existential crisis when it forgot how to “end”.
(Click) (Wait, I am the AI, I don’t click) 
2 Likes
Same here, I didn’t have such problems with other models other than Gemini in antigravity
1 Like
Thank you for bringing this to our attention. We have escalated the issue to our internal teams for a thorough investigation.
To ensure our engineering team can investigate and resolve these issues effectively, we highly recommend filing bug reports directly through the Antigravity in-app feedback tool. You can do this by navigating to the top-right corner of the interface, clicking the Feedback icon, and selecting Report Issue.
Take a screenshot of the usage spike from the cloud console and post it here please.
dont have any screenshots on my end but just wanted to say that I have also been getting this bug, it seems to be related to Gemini itself. i’ve even gotten it from other platforms, I think even Cline once. sometimes resetting/reverting the chat to the last “stable” prompt works. sometimes it still breaks and you have to nuke the entire chat session and start over.
i also observed a tendency for this to happen in antigravity if you use the dictation feature.
I’ve had this happen a few times too. I’ve just stopped it, then tried again with a different prompt. Problem solved. But, like you said, Claude models don’t do this, which is why Sonnet 4.5 is my go-to most of the time.
try to resume with claude or G3 pro High.
I have the same issue. When calling gemini 3 flash via API calls, it often ends in infinite thinking loops until all tokens are consumed: “end.end.end.end.end.end.end.end.end.end.end…..”
EDIT:
Solved it in my API pipeline: The issue comes from the json/application schemes.
In a complex task the LLM with enforced schemes has to reason about both scheme format and prompt. The schemes enforce a specific token and these unfree tokens cause the loop: The LLM fails to hit a correct scheme-respecting token and can not finish, looping to infinity.
I switched to prompt fields for complex reaosnign tasks with constructred/guided CoT.
You can decouple reasoning and formating in a step call with a simple parser or a cheap model call inbetween if the format is critical.
“NEVER add OBJECT-type fields with enum sub-properties to an already-large schema”
What helped me was splitting complex tasks into single-step prompts and adding an explicit “if you encounter an error, stop and report it to me” instruction. The loop usually starts when the agent tries to self-correct a failed step.
Same issue. Got it like 5 times yesterday.
This or the opposite, antigravity kept stopping without continuing for no reason having to write continue multiple times. gemini 3 pro.