I am a subscribed Ultra user. I’m writing to report some severe stability issues that are hindering the workflow, along with a specific anomaly I observed that might help you debug the backend logic.
1. The “Agent Execution Terminated” Loop (The Core Issue) I am constantly encountering the “Agent execution terminated due to error” message when using Claude 4.5 Opus.
Symptoms: Every single step requires a “Retry”.
Contrast: This issue rarely happens when I switch to Gemini models.
Diagnosis: This strongly suggests a backend timeout configuration flaw. Claude 4.5 Opus (and high-reasoning models) has a longer Time-To-First-Token (TTFT) and generation time compared to Gemini. It seems your backend kills the process before the model finishes its CoT (Chain of Thought), treating it as a “hung” process.
2. A Strange Anomaly (Potential Debugging Clue) Here is a very counter-intuitive observation I made: When the Context Window is extremely large and the generation speed is visibly slow, the error actually triggers LESS often.
Hypothesis: This implies your system might have a different timeout policy for heavy-payload requests, or the continuous stream of tokens from a large context somehow keeps the connection alive (heartbeat) better than a “thinking” state where no tokens are generated for a while. This inconsistency confirms the timeout logic is brittle.
3. Feedback on the Quota System (Community Pain Point) While I personally benefit from the Ultra quota, I’ve noticed significant frustration in the developer community regarding the rigid 7-day refresh cycle. The logic seems prone to glitches where the 7-day timer resets incorrectly, locking users out. This unpredictability is damaging the user experience for non-Ultra developers.
Proposed Technical Solutions:
Dynamic Timeouts: Adjust the backend timeout threshold dynamically based on the selected model. Claude Opus should have a significantly higher tolerance than Gemini Flash.
Heartbeat/Keep-Alive: Implement a “thinking” signal or heartbeat mechanism. If the model is processing but hasn’t output a token yet, the agent should not be terminated.
Soft Fails: Instead of terminating the whole agent, allow the frontend to wait longer if the backend reports “processing”.
I enjoy the tool when it works, but I currently have to rely on my own Rust-based workaround because the platform’s stability with SOTA models is unreliable.
Sorry for delay. No, we don’t know how long it will take or even if they will fix it. I talked to google customer support person and they didn’t know either. They said the system handles quotas so unless something in antigravity changes nobody can stop these rate limits.
The screenshot just shows that according to the antigravity logs I’m not actually logged in even though the antigravity IDE indicates that I am logged in. This same situation has affected many people.
Same issue here. Renders the app basically unusable right now. It would be easier and more transparent to offer a possibility to pay for the credits on the go instead of guessing. Now this days-long freeze took place practically without doing anything with Cloude.
You’ll need to reinstall Antigravity for the “Agent execution terminated due to error” error to go away. I Googled it and did it and now it is not happening for me
It’s all very clear that Google deceived us. We paid money and subscribed.It’s all very clear that Google deceived us. We paid money and subscribed. I want my data allowance, which renews every 5 hours, back.
Everything was working fine till i was blocked till the end of the week. I waited and when the week ended they changed the date to another week. Now it gave me 40% which I used in 2 hours and I have to wait another week. And no I didnt do much in those 2 hours of using it. Before there was the 5 hour wait but now is worst than a free account and i paid yearly pro ai.
This morning I went to do a quick fix to fix a bug, and I already got the 10-hour limit on the Germini model, and the cloud had already reached the 7-day limit.
It’s complicated, I sent an email, tried to contact support, and nothing, no solution!
Same with me…i dont know its like not subscribe google ai pro…so until its clear this issue or new term so iam make a stop continue subscribe if same like free
This has started happening after antigravity v1.15.8 update to everyone.
I updated yesterday and things wend bad, before this update everything was perfect.
I understand that google change limits for many models, but the AI Pro subscription was not affected from my understand according to Google’s intimation. Specially when even Anthropic’s reset time is lower than this.
I mean I just updated from 1.14.2 to 1.15.8 yesterday and started facing these issues. Everything worked properly before then.
I am talking about fresh accounts. My brother doesn’t codes, but pays for AI Pro subscription, so he logged into my system to test, and same results (he has never coded or used a single model request before on this account)
I am running macOS (i dont know if this is also on windoews/linux).
I downgraded using the normal installer from the antigravity website and going to previous versions.
I mean I have been using 1.14.2 ever since it released until yesterday.
Unfortunately, did reinstalling antigravity work? I’m not sure; it’s just that the ‘Agent execution terminated due to error’ issue decreased from occurring every step to once every three steps. My system is Ubuntu, and the output is still: UNAVAILABLE (code 503): No capacity available for model claude-opus-4-5-thinking on the server: UNAVAILABLE (code 503): No capacity available for model claude-opus-4-5-thinking on the server.
Same here. At the beginning of the pro account in December it was not like that. Go a lockout until 3/2/26 we can’t see anywhere the current quota in antigravity and how it works
