Google dropped Antigravity 2.0 at I/O 2026. I installed it minutes after the stream ended. Took me three attempts, but I got it running. Here’s what I found — and why I’m not uninstalling VS Code just yet 🔥
The installation gauntlet
I hit “download” the moment the keynote ended. First attempt? Some DLL error. Second? Login refused to cooperate. Third time’s the charm — it finally launched.
Not the smoothest onboarding, but hey, it’s day zero. I’ve seen worse.
Where’s the IDE?
I open it up and… where’s the code editor? I’m clicking around, looking for the familiar split-pane view, the file tree, anything. Nothing.
Google calls this an “agent-first” development platform. What that actually means: you get a chat window to talk to the agent, and that’s basically it. No classic editor. No file browser. Just you and the agent, staring at each other.
On one hand, I get the vision. On the other — I’m a developer. I need to see my code, jump between files, make quick edits. I’m not ready to hand over the keyboard and just vibe in a chat window. So yeah, I’m staying with VS Code.
Gemini 3.5 Flash — from pigeon to actually useful
Here’s the thing: literally ten minutes before the I/O stream started, I was wrestling with the old Gemini model to finish a migration I was in the middle of. It was painful. The thing was a blockhead. It felt like trying to code with a pigeon — you point at something, it stares blankly, then does something completely unrelated.
Then the stream ended, I switched to Gemini 3.5 Flash, and… it actually got stuff done. Finished the migration. Didn’t hallucinate halfway through. It’s fast, it understands context, and it just works. The difference is honestly night and day.
I’ll be testing it more, but first impressions? Solid. There’s real progress here.
The permissions harness is still annoying
The one thing that drives me nuts — and drove me nuts in the previous Antigravity too — is the permissions system. They call it the “agent harness.”
Here’s how it goes: I give it permission to do something in a project. A minute later, it asks again. For the exact same thing. I say yes. It asks again. Eventually I give it blanket permission, and suddenly it remembers and works fine.
I had literally the same problems in the previous version. Google, if you’re reading this: if I said yes three times in the last five minutes, maybe just trust me?
No custom models
I looked around — no option to hook up custom models. No OpenRouter, no DeepSeek. Maybe it’s buried somewhere, maybe it’s coming later, but for now you’re locked into Google’s models.
So, is it worth it?
Antigravity 2.0 looks interesting, and Gemini 3.5 Flash is genuinely fast and capable. But upgrade the entire tool just to get access to a better model? I don’t know. The lack of a proper code editor is the dealbreaker for me.
For now, I’ll keep Antigravity around for testing and certain agentic tasks, but VS Code isn’t going anywhere. Flash is great — but I can use great models in tools that actually let me write code too.