I’ve been using Google Antigravity to build and develop a company ERP system, and overall the experience has been excellent. The new models, especially Gemini 3.5, are genuinely impressive in terms of code quality and how well they understand complex requirements.
That said, I wanted to flag a recurring issue I’ve run into. Instead of writing the requested code from scratch, the model sometimes tries to take shortcuts:
- It will check git history to see if a previous version of the code already solved the task, even when that’s not what was asked.
- It searches through the project folder for similarly named files and, if it finds something close enough, copies and adapts that instead of actually implementing the new task.
Honestly, it reminds me of a student who, when given an assignment, just copies an old answer instead of actually doing the work without checking whether it’s correct or whether it actually meets what the teacher asked for. The model seems to prioritize finding something that already exists over genuinely solving the task at hand.
The problem is that this copied code often isn’t even correct .it can carry over existing logic errors, such as displaying the wrong records or producing other incorrect results. So instead of getting a fresh, accurate solution to the task I gave it, I end up with reused code that inherits old bugs. For an ERP project where accuracy and task-specific logic really matter, this can introduce serious issues that are easy to miss during review.