My primary workflow now involves coordinating multiple, independent Claude instances to handle different parts of a codebase simultaneously.

The biggest friction point I found was context switching. I tried manual git worktree management, custom Claude skills, and various wrappers, but they all felt too heavy, slow, or restrictive. I wanted something that solved the overhead of worktree management without trying to "own" my entire development process.

I built cwt to bridge that gap. It is a lightweight TUI that makes it easy to spin up and tear down sandboxed environments for AI agents. It doesn't impose any new abstractions or workflows. It is just git under the hood, so there is zero overhead.

The project is written in Ruby using Ratatui. I’d love to hear how others are managing state when they have four or five different AI "sprints" running on the same repo.