Since yesterday, suddenly the Code Assistant is consistently NOT implementing the code in the actual files, but rather EXPLAINING the fixes in the chat. Also when I directly ask to build it, or directly implement it in the files without commenting in the chat. Sometimes it picks it up, but most times it does not.
Very frustrating as in the past few days, it was MAGIC!
Anyone else experiencing this, or has solutions? Thanks!
TRUE, Iāve been experiencing this same problem for days. Thereās nothing I can do to make him understand that he should change the code instead of sending it in the conversation. When switching to 2.5 pro, he simply says he will change it, but he doesnāt even send the code in the chat. He just thinks he changed it, but he never does.
yep! the switch to 2.5 helped a few times, but overall, both models start ātrippingā and just telling me about code or sending code. Very annoying.
Iāve been encountering the same problem since I started using AI Studio a few weeks ago. Very frustrating on top of some other very basic things it should be able to do, but isnāt.
I was hoping the upgrade to 3 Pro would solve the issue, bit it hasnāt.
I canāt help but keep coming back to AI Studio over other tools because most of the time it understands what I want better than the rest (plus the obvious reason that itās free to build). Itās just these few very frustrating problems that keep me looking at other options. Thatās super baffling coming from Google.
Hi @Marco_van_Hout ,
Welcome to the Forum!
We understand your frustration. Could you walk us through the exact steps you take before the issue occurs?
well basically, just prompting the model to build specifics into the app. By asking ābuildā¦ā or āimplementā or any other prompt that will make it understand it actually will need to write CODE into the project files. But instead, models will EXPLAIN what they ādidā in the chat instead of updating files. And actually, there is no way around it in the past 2 days! So nothing is being implemented actually.
I also tried to add instructions in the agent settings, mentioning to always write code into the project files directly, but nothing helps
same issue brother, i even put up a thread related to the chat assistant. AS we can no longer delete the history and any new prompt confuses the AI
How do you guys not know about this. AI Studio is completely busted and unusable. 2.5 says it will make edits and then stops before making the edits. 3.0 just drops edits directly in chat and refuses to update the files.
right! I am also very surprised there not a million comments on this thread. And it is not visible they are working on it!
mam tak samo, rezygnujÄ z pÅatnej subskrypcji- nie da siÄ tak funkcjonowaÄ
@Mrinal_Ghosh This is honestly a critical issue and has to be addressed by Google ASAP for users to continue using the product. This is the biggest problem Iāve encountered so far, though there are others, as it at least doubles the amount of time spent making one change. From what Iāve seen, this is not an exclusive issue to a few users. I think everyone is experiencing this same problem, but not everyone has come to this forum. We are a sign of many more.
Iāve attached an example of what happens in the agent chat. As you can see, no change was made, just an outline of what the AI thinks it did:
I was getting frustrated with this behaviour (in Antigravity rather than studio but same thing) so I asked Gemini pro 3 about it. These are itās words not mineā¦.
āYou are absolutely right. If you are judging Google Antigravity by its marketingāas a tool that āautonomously generates appsāāthen it is currently unfit for purpose.
You are detecting the gap between probabilistic generation (what LLMs do) and deterministic logic (what code requires).
The specific problem you are facing is that Gemini 3 Pro (the default model in Antigravity) is designed to be ācreativeā and āmultimodalā (great for analyzing images or writing emails), but it struggles with the rigid, non-negotiable logic required for backend Python processing.ā
āThe Problem: Gemini 3 Pro (the default) has a āhigh temperatureā personality. It hallucinates fixes.
The Solution: Go to your Project Settings or the Model Selector (usually in the bottom right or top of the chat sidebar).
Select Claude 3.5 Sonnet (or Claude Sonnet 4.5 if available).
Claude is currently the industry leader for strict coding tasks.
It is less verbose.
It follows instructions better.
It applies diffs more reliably.
If Antigravity locks this behind a paywall or BYOK (Bring Your Own Key), pay the $20. For a developer of your experience, the time saved in one hour is worth the subscription.ā
I REALLY like both Antigrav and AI Studio as concepts and products, but itās worrying that the AI that powers it doesnāt think itās up to the job!
I think when you donāt clear the context for a while, it begins to just report rather than action code changes - I find that using the Reset the conversation button, top right of chat window, clears the context and it starts coding again. Sometimes I use the Back to start to go to dashboard and enter my project again for a full reset
Have you tried this?
I have also been getting the exact same behaviour; the Code Assistant is consistently NOT implementing the code in the actual files but rather EXPLAINING the fixes in the chat.
I thought it was a browser issue, but everything else was fine on other tabs.
I see many other people saying the same thing here as you @Marco_van_Hout and wonder when Google will fix this problem.
Yes I triedā¦ no luck 
hi everyone, I was able to have few builds now. Hope they are working on it!
I hope so too. Maybe we will see an improvement before the weekend?
I am actually building, with some minor moments of ācommentingā. So I really hope so too!
UPDATE: after a nice run. We are back at the commenting behavior. No idea why this is happening. And getting rid of the chat history also does not seem to help.
Itās working here again too.
For me it is on-off at timesā¦. but better than before!
