Carbon has been an extremely impressive model so far, having used it for a grand total of four or five hours, but Antigravity feels primitive compared to Claude Code. Run it with --mode auto and it still pesters you to death. Run it with --dangerously-skip-permissions, like I'm doing now, and who knows what will happen. Guess I'll find out.

When I first installed it, it took 10 minutes of arguing before I convinced it to add the context percentage and the current working directory to the status line at bottom. The fact that harnesses still don't present these insanely-vital statistics by default is puzzling, but Claude made it a lot easier to add.

So my first impression of Antigravity was to say the least not great. When it comes to actually orchestrating model and agent activity, though, it seems to be doing what you'd expect it to do.