I am constantly running into what appears to be endless loops? I have been leaving these running for ages but they never stop. I am assuming this is not right? I also hope this isn’t chewing through my quota!
Oh, it definitely is.
Hi @Matt_Hall,
What you’re seeing is a tool-use loop, and yes, it will rapidly chew through your quota since each loop iteration resends the entire accumulated conversation history.
To prevent unnecessary quota usage, I recommend hitting Cancel as soon as you notice the agent stuck in a repetitive loop. From there, it’s best to start a fresh session rather than trying to steer the existing thread, as the repeated calls in the history can bias the model to keep looping.
Flash models benefit significantly from explicit boundaries and fallback instructions. For example:
“Inspect useHoldingsStats.ts once. If you cannot identify the issue, stop and ask me directly instead of re-inspecting or guessing.”
If you’re tackling complex debugging tasks, you can also consider switching to Gemini Pro or a model with deeper reasoning/thinking enabled.
Lastly, please report this via Settings > Provide Feedback > Report Issue (bottom-left in Antigravity 2.0). Submitting your session diagnostics helps the team improve client-side loop detection and agent safeguards.
Thanks @chunduriv, I will follow these steps in future. I assume putting something in agents.md might help too? Rather annoying it has burnt through about 10% of my weekly quota, as it was doing this multiple times.
@chunduriv just to let you know though, I have been using it for a few more hours this morning and even with prompts, changes to the agents md and starting new sessions, I frequently get in a position where it is doing this still and/or getting stuck in other analysis “loops” where it never seems to stop analysing various files. This is completely new behaviour I have never seen in any other model this year.
The other I saw Gemini taking forever so I did:
/btw Are you stuck in a loop?
It then said it was almost finished and wrapped things up pretty quickly. Maybe it needed a nudge.
I got the same issue too; the problem is that I’ve been trying the remote control mode. So this is not something we can “catch” and give it a nudge. Since we are trusting things to work, when we get back, it will have finished the job. probably the better course of action is to use 3.7 until a new launch next month fixes this