In my previous post about Pi, I wrote about wanting a stable coding harness while the models underneath it keep changing. But one agent in one terminal is not the whole workflow. I increasingly have several agents working at once, sometimes in the same repository. The surprising part is that I manage that with a simpler stack: a terminal, Git worktrees, and a terminal multiplexer built for agents. My current setup is mostly Ghostty, Git worktrees, Herdr, and hunk.