I built a tmux‑style split‑pane app to run multiple Claude Code agents during an interview. Each pane = a Claude session + its own git worktree. You get real‑time token usage and visible merge conflicts. Repo: https://github.com/MatchaOnMuffins/orchestrator
Why this matters: parallel agents are powerful but messy. Without isolation you get branch conflicts, context bleed, and weird commits. Pane per agent + worktrees keeps state scoped and context visible at a glance.
Practical bits worth copying: one git worktree per agent, token usage telemetry per session, surface merge conflicts before commits, tmux keybindings for fast navigation. Tradeoffs: BYOK/local keys and mac packaging/signing friction.
Takeaway: run agents like separate services — isolate state, track costs, and show conflicts early. Otherwise you’ll spend more time resolving git than reviewing output. How are you isolating multi‑agent workflows today?
Top comments (0)