It’s a good feature. However I am not convinced that the terminal/ssh abstraction will remain optimal when you orchestrate a swarm of workers across multiple machines.

Running via ssh is great when your machines are pets that have local state. I don’t find it to be that useful when I have more than two machines that I’m load balancing agent sessions across.

Just for example, a nice feature from the Codex app; if you edit a markdown file on your remote, and click on the link to the edited .md file, it just opens the file in the local app, and you can view the rendered form too. If I’m working on my local machine this isn’t a win vs. opening the file in Obsidian / IDE. But it’s way easier than manually viewing the file over SSH (unless you just want to look at the raw text version). Similar applies to html files.

I think it’s more likely the Fly Sprites or Claude/Codex sandbox model will win as we move to swarms. If I’m not watching the actual harness output, there’s not a great reason to attach a pty.