Guide
tmux vs. git worktrees — what actually keeps parallel agents apart?
Run five coding agents on one repo and two problems appear immediately: their edits collide, and their processes die the moment your terminal does. Orchestrators answer with two different primitives — git worktrees for the code, tmux sessions for the processes. They're often mentioned in the same breath, but they solve different halves of the problem — and plenty of apps use both.
Git worktrees: isolating the code
A worktree is a second (third, fifth…) checkout of the same repository, each on its own branch. The orchestrator creates one per agent, so every agent edits its own copy and ships its own branch — merge when you're happy. It's the de-facto standard across desktop apps like Conductor and web dashboards like Vibe Kanban. What worktrees don't do: keep an agent alive when its terminal closes.
tmux: isolating the process
tmux is a terminal multiplexer: it runs shells in detached sessions that survive SSH drops, closed windows and laptop lids. A tmux-based orchestrator — Claude Squad, dmux, Agent of Empires, jat — parks each agent in its own session and gives you a switchboard. The killer feature is re-attachability: start agents on a remote server, disconnect, reconnect from anywhere, sessions intact. That's why tmux dominates terminal-UI orchestrators and remote-server workflows.
Not either/or
The best setups combine them: worktrees so agents don't overwrite each other, tmux so sessions outlive your connection. Many tmux-based tools create a worktree per session automatically. When you see both listed on an app's spec sheet, that's the pattern.
| Model | Isolates | Strength | Limit |
|---|---|---|---|
| Git worktrees | Code / branches | Zero-conflict parallel edits, easy review | Process dies with the terminal |
| tmux sessions | Processes | Survives disconnects, remote-friendly | No code isolation by itself |
| Containers | Whole environment | Safety sandbox for auto-approved agents | Heavier setup, Docker required |
| Cloud VMs | Whole machine | Runs while your laptop sleeps | Code leaves your hardware |
How to choose
- → Supervised work on one machine? Worktrees are all you need — pick any desktop or web orchestrator.
- → Agents on a remote server, you on a laptop? tmux-based — filter the directory for tmux.
- → Unattended agents with broad permissions? Containers or a remote-server setup — blast radius matters more than convenience.
FAQ
What does tmux actually do in an AI agent orchestrator?
tmux keeps each agent in a detached terminal session that survives SSH disconnects, closed windows and laptop sleep. The orchestrator spawns one tmux session per agent and gives you a UI to list, attach and manage them — the agents keep running whether or not you are looking.
Do git worktrees and tmux solve the same problem?
No — they are complementary. Git worktrees isolate the CODE: each agent gets its own checkout on its own branch, so parallel edits never collide. tmux isolates the PROCESS: each agent keeps running independently of your terminal. Many orchestrators use both at once.
Do I need to know tmux to use a tmux-based orchestrator?
Mostly no. Tools like Claude Squad or dmux drive tmux for you behind a friendly TUI. Knowing tmux basics (detach, attach, kill-session) helps when you want to poke at a session directly, but it is not required day to day.
When are containers or cloud VMs the better isolation choice?
Containers add a safety boundary: an agent running with auto-approved shell access can only damage its sandbox, not your machine. Cloud VMs go further and move the whole workload off your hardware. Choose them when you run agents unattended or with broad permissions; worktrees and tmux are lighter when you supervise the work.
Compare the tools that use them
Filter every orchestrator by tmux, isolation model and more.
Browse the directory →Last updated July 2026.