I’m not going to go into details on my background here.
But I am going to talk about my experience with Antigravity since I started using it a few months back.
MY EXPERIENCE:
Wanted to work faster with agents (agentic development),
I did a deep dive into the documentation and many expert videos later.
Finaly gave this ‘vibe coding’ a try…
Got very frustrated with the agent chat view UX on day 1.
Made changes to the rules of the agent.
Created a new expert model mode to change models by tasks.
Added TTS to the agent, nicknamed it ‘GRAV’ for focused recaps after every prompt.
Made it even work outside Antigravity IDE to make it work faster (Logs and Batch files), because it asked for to mush confirmations for what I am working on or asking it to do (No allow always option).
Created a off-IDE white- and blacklist for the commands.
Changed the settings I needed to work faster.
Created a handoff procedure to make the model communicate to other models (including offline models), similar to A2A.
Did more research into why it forgot simple top rules, like the use of the TTS I made for it and had to restart every hour to cure it and to make it follow the rules again.
The previous flash model almost gave me AI psychosis.
all and all, it was working fine, I was making real progress.
I finally got accustomed to it…
Getting this out of the way, if I am using it, is because I like it, now moving to the main points I want to make.
RANT:
Seen the new 2.0 gpt-like UI, did not like it (Not for me), many of the new features I have already made myself, had to restart the IDE.
Previous settings are gone (I see I am not the only one, I will try the fixes I see here in the group),
Now the new models talk and over think to mush (even more then the previous flash model), it does less actual coding, uses more technical jargon, over explains, over analyzes, over thinks, does more scans, even if I stop it and tell it where to look for what it is looking for (path and line number), forgets rules faster, goes off on tangents, does not understand like the other model (somewhat did), it reaches quota limits faster then ever (5 to 10 times faster), got very curious about all this, seen the price hike by 20 to 60 times the previous model (Had to confirm it with Gemini Web, it did confirm it).
I now even know not to use STT (Voice) for the agent, before it had a 10x token cost for each minute, the audio is not optimized or compressed, before sending it (It goes out RAW, also did confirmed it with Gemini Web), now I made a layer to make it transcribe the audio to text…
Even had to take time to make a personal prompt optimizer to consume less tokens, changed rules to give no/some/all feedback, recaps, no logs, less stability monitoring depending on the current quota limit levels, to minimize tokens sent and tokens consumed in the responses.
Had to view many videos about all these new changes, I see many people are on the same boat as me.
NOTE: My Youtube “negative” observation comments are not showing for anybody or most (personal friends and VPN PROXY web scraping did confirm this), I hope this does not happen here in this group.
Many people are already jumping ship, but this is a development ecosystem wide issue, not just about this IDE, Gemini or Google, wherever anybody migrates to, sooner or later this same issue will happen.
Feel free to comment, share your similar experiences and also to correct me if some of the info is incorrect (Even Gemini Web hallucinates and pulls facts from nowhere), I know how to take critics (I am an old-school 50yo dev with no EGO, not a rockstar dev or ever worked at any of the FANG companies).