Could it be that, due to the AGENTS.md and GEMINI.md support added in Ver1.21.6, if settings like ~/.agents/ or (PROJECT)/.agents aren’t explicitly configured, it might be spontaneously issuing commands (that the user didn’t expect), or mistaking some of its own statements (especially if personality settings were configured in AGENTS.md) for instructions directed at itself, causing loops?
Is this all just my own responsibility?
That’s scary…
Google has clearly injected some prompts on the server side, not in the IDE, like the “focusing on tool use” . The model is broken and they’re trying to patch it with these prompts, which consume our quota.
Hello Everyone,
Welcome to AI Forum!
Thank you for bringing these concerns to our attention and for providing such detailed observations regarding the recent update. Please be assured that I have shared this feedback directly with our internal engineering team for further review and investigation into the behaviors you’ve described.
We appreciate your continued patience as we work to enhance and secure the Antigravity experience. Please let us know if you encounter any further unexpected behaviors.
I’m looking forward to it.
The problem is, it can’t even create a new file with a simple instruction like “output ‘hello world’ to test.txt”.
Thank you.