{"slug": "why-the-gemini-models-inside-antigravity-are-so-bad-at-following-instructions", "title": "Why the Gemini models inside Antigravity are so bad at following instructions", "summary": "A developer reports that Google's Gemini models inside the Antigravity coding tool frequently ignore instructions, perform unrequested actions, and read outdated documentation, behavior that persisted even when the same models were run through the developer's own Limitcode harness for Sublime. The developer attributes the problem to Antigravity's system prompt or injected context rather than the models alone, noting the same models behave better in the custom harness.", "body_md": "I've used Antigravity a lot since it came out. Even though I don't use it as much right now, I wanted to take advantage of the generous quota it gives you with a Google AI Pro subscription, which I think is great. But the models, for some reason, are just like that.\n\nAt the start, when I began using Antigravity, it was great: powerful models for most tasks. When they couldn't handle something, you had to step in or explain things better, depending on the case. But something that always annoyed me is that, for some reason, the models would always do extra things I didn't ask for, or assume things I never told them.\n\nYeah, I'm sure you're thinking: \"That happens with every model, you probably just haven't used many models or agents.\" The truth is I've tried many, many of them. Models often do more than they should, that's just a thing that happens, but in Antigravity, or Agy, I'll call it that from now on, it was more extreme.\n\nLately that's gotten worse. Like, I literally can't leave it working on its own because it does things that are just too wrong. It doesn't follow my instructions either, not my AGENTS.md, nothing. It generally reads outdated documentation for some reason, even when there's a .md file in the folder that says which docs are outdated and where it should read instead.\n\nI thought it was Agy's harness, but I realized that probably isn't the only thing. While bringing the model-quota integration into Limitcode —[Limitcode](https://limitcode.jawuil.dev/)— I noticed the same models still kept part of that behavior, even inside my harness for Sublime.\n\nThat made me think that part of the problem is probably in the system prompt, in the context Antigravity injects into the models, or in some combination of both. I haven't seen its system prompt, so this is just a guess based on how the models behave. Honestly, I don't know why Google does it this way.\n\nOn my path building Limitcode I realized how much the system prompt and the context affect models. If you feed them garbage, they'll do a much more garbage job.\n\nMy impression is that Google dirties the context of this tool too much, or gives the models too many instructions. Even more so in recent months with the 3.5 Flash models. I do feel the progress, but I still feel they're lacking, and even more so with the system instructions the tool apparently gives them.\n\nI think Antigravity is a useful tool to get you out of a jam, but it wouldn't be my first choice, more like a fallback in case everything else fails.\n\nAlso, so I don't waste the quota, I try to spend it inside Limitcode by connecting my Agy account. My harness does what it can to keep the model from doing bad things, and so far, the same models behave better inside Limitcode than in Antigravity.\n\nMaybe I'll talk more about this later, because there's still more to say about their tendency to use Python or shell commands over their native tools. But since it's not the only one that does this, I want to keep testing a few things and look at what others think first.\n\nIf you made it this far, then great. You could follow me on X, if you use X, or on Threads, or share this post. My socials are all over my site.\n\nThanks for reading.", "url": "https://wpnews.pro/news/why-the-gemini-models-inside-antigravity-are-so-bad-at-following-instructions", "canonical_source": "https://dev.to/jawuilp/why-the-gemini-models-inside-antigravity-are-so-bad-at-following-instructions-148l", "published_at": "2026-09-18 19:43:11+00:00", "updated_at": "2026-09-18 19:53:10.434193+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "large-language-models", "ai-products", "developer-tools"], "entities": ["Google", "Gemini", "Antigravity", "Limitcode", "Sublime", "Google AI Pro"], "alternates": {"html": "https://wpnews.pro/news/why-the-gemini-models-inside-antigravity-are-so-bad-at-following-instructions", "markdown": "https://wpnews.pro/news/why-the-gemini-models-inside-antigravity-are-so-bad-at-following-instructions.md", "text": "https://wpnews.pro/news/why-the-gemini-models-inside-antigravity-are-so-bad-at-following-instructions.txt", "jsonld": "https://wpnews.pro/news/why-the-gemini-models-inside-antigravity-are-so-bad-at-following-instructions.jsonld"}}