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.
I've resorted to running the full harness apps in a macos VM with full permissions and connect from host over SSH*. Code is mounted as a shared directory. It's been working well so far. I plan to write a blog and publish code on how to setup the VM with the tools if folks are interested.
* - In some cases I use the harness specific method to connect e.g. Hermes agent and OpenCode GUI app on host
Haven't used Antigravity (mainly Claude and Codex). Is it worth the switch? I've been pretty annoyed by Claudes usage opaqueness.
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.
I've resorted to running the full harness apps in a macos VM with full permissions and connect from host over SSH*. Code is mounted as a shared directory. It's been working well so far. I plan to write a blog and publish code on how to setup the VM with the tools if folks are interested.
* - In some cases I use the harness specific method to connect e.g. Hermes agent and OpenCode GUI app on host