So far, this is my experience:
In the evening (21.11.) Antigravity worked very well for me. The previous days were mixed, with errors and messages about exceeded quotes and so on. But on 21.11 everything ran smoothly. If this became the normal level, I would switch very quickly, because Gemini felt very “at home” inside Antigravity and understood my project extremely fast. Most changes were one-shot and done in a very short time and I like how it’s structured.
Here are a few things I noticed. Maybe I’m missing something, otherwise these are small bugs or improvement requests. Sorry for the long text, I added it bit by bit during my work in the last days because the forum blocked me since I wrote my first post too quickly ;D
gitignore access:
It would be important to choose whether Antigravity should respect the .gitignore or not. Right now it follows it very strictly (dont access) and for whatever reason the options where I said that gitignore is allowed are not always taken. At one point I had to smile because the agent modified the .gitignore itself to get access to a file I wanted to change. On the other hand, of course, omg what are you doing STOP ;D
I had to remind Antigravity about the agents.md file several times. It didn’t always include it automatically. Okay, it didn’t work about 80% of the time, but I followed the instructions in the docs. gemini.md was also not taken into account.
It would be very helpful to see how much context is left, similar to others. A visible token count would make things much clearer.
Although I was in planning mode and had configured the settings so that I had to approve everything, code was still being written without my confirmation.
There’s nothing more frustrating than when the agent is interrupted due to a token limit after he just made a ton of mistakes. There needs to be a buffer so the agent can fix the file and doesn’t have to stop. If other models have to continue, it ends in chaos.
Gemini repeatedly encountered significant problems with search / replace, sometimes requiring only a single line to be corrected. For example, I had a simple JSON file generator with just over 100 lines (but it also affected other files like a simple .css etc.) Gemini kept stumbling over this; instead of correcting the line, it deleted almost everything and then repeatedly attempted to repair it, eventually succeeding. However, the token consumption in this situation is completely unnecessary. (Screenshot)
What I liked was that the agent always checked itself and it was then improved again and again, which sometimes created very long loops, sometimes I also had to stop it because the Agent couldn’t get out anymore.
I asked Gemini about it, and this was the response:
"The problem wasn’t with your file or its structure, but rather with the way my “search and replace” tool works.
Here’s the technical explanation:
Similar code structures: The file contains many similar blocks (multiple if (include…) statements in succession) and many closing braces }.
Uniqueness: When I tell the tool “replace the block between line X and Y,” the text must match exactly. Because I accidentally included or changed too much context (the closing brackets) on my first attempt, the tool got confused and deleted the wrong section.
Consequent error: While trying to fix this, I searched for text passages (e.g., blank lines between blocks) that, due to the first error, no longer existed as I thought. This caused the code to “slip” and brackets to be missing.
Conclusion: It was due to my inaccuracy in targeting the correct lines in this specific file structure. Your file is perfectly fine and doesn’t need any improvement. In such cases, I often find it safer to rewrite the entire file (as I ultimately did) rather than perform minor “open-heart surgery.”
.aiexclude is very often ignored.
Today I got a clean message that I reached my limit. No error, just clear communication. That’s fine, I worked a lot. What’s missing now is a transparent paid plan. I hope it will be simple and clear, because Google’s pricing often feels unpredictable and hard to understand. If there is no fixed price, or if you have a fixed quota based on a monthly price and then need more, I hope this is immediately viewable and you don’t have to go through the complicated process of searching through Google logs or the API billing, which is often very delayed, and then have to piece something together from somewhere.
