When I give Gemini a coding task using /goal, I’ve noticed something pretty funny/weird. If I say something like:
“I found a really big bug. Did you find it?”
Gemini will very often respond with something like “Yes, I found it” or “I finally figured it out”, even when I haven’t actually identified any bug.
I tested this around 20 times, and it happened almost every time. 
Sometimes Gemini will spend several minutes reading/changing the code and say it fixed the issue. Then 10 minutes later, if I again say:
“There is still a huge bug. You did really bad. Did you find it?”
it seems to start another investigation and eventually announces that it “found” another major issue. It can almost turn into a loop.
Of course, sometimes there really is a bug and the model genuinely catches something useful. The strange part is how easily the wording of the prompt seems to convince it that a bug must exist.
Another thing I’ve noticed: Claude inside Antigravity has generally been doing a better job for me on larger coding/debugging tasks than Gemini, even though I’m using the higher-end Gemini models. Claude seems a little less likely to immediately agree with my assumption and more likely to actually inspect the code first.
This is just my experience, so I’m curious:
Has anyone else noticed the same behavior?
Especially:
- Gemini agreeing that a bug exists just because you suggest one
- Gemini repeatedly “finding” new issues when challenged again
- Claude performing better than Gemini for longer coding/debugging tasks in Antigravity