The key for me has been that they are useful when I'm working on a new product or a very very large feature. After I have planned with claude for a few hours to refine architecture and edge cases, my implementation plans are rather large but broken into phases.

I usually ask it if any steps are good for parallel work, and I kick off multiple agents at that point. Each is assigned a phase. They have to work on different work trees in my flow and I've rather disliked it and I wish there was a better way than worktrees at the moment, but I guess people are busy building those tools, right?

Hell, I'm writing a small tool to let me share a Claude session with my team from their own terminals, so that any of us can take prompt control while we work through plannings and implementation. I'm sure many others are finding these needs!

I often have both Claude-Code and Codex-CLI running in separate Tmux panes. And I have them use my Tmux-cli tool to verify, validate, consult or delegate to each other:

https://github.com/pchalasani/claude-code-tools/tree/main

I wrote up some tips on that here: https://cmux.io/prompting-tips.html — most of it is generally applicable outside of `cmux`