How to Run Multiple Claude Code Sessions in Parallel (Without Losing Track)

One Claude Code session is a tool. Five running at once is a team — one refactoring the auth module, one writing tests, one grinding through a migration, one on docs, one chasing a bug. Parallel sessions are the real productivity unlock of agentic coding. The catch: somewhere around the fourth terminal you stop being a developer and start being an air-traffic controller who's lost the radar.

Here's how to actually run Claude Code in parallel — and, harder, how to keep track once you do.

Running them in parallel

The mechanics are easy; the discipline is the trick.

• A session per terminal. iTerm/tmux split panes, or separate tabs. Run claude in each, one task each.
• A git worktree per session — this is the real move. Multiple agents editing the same working tree stomp on each other. git worktree add ../repo-tests tests-branch gives each session its own checkout of the same repo, so they can all work at once without collisions. One worktree per task, one session per worktree.
• Scope each one narrowly. Parallelism works when tasks are independent. "Extract the token-refresh helper" and "write the parser tests" don't fight; "refactor everything" and "also refactor everything" do.
• Resume, don't restart. claude --continue picks a session back up where you left it, so you can tab away and return.

Do this and you genuinely get multiplied throughput on independent work.

Then you lose the thread

The failure mode isn't technical — it's attention. At five-plus sessions:

• You can't remember which terminal is the migration and which is the tests.
• One finished ten minutes ago and has been sitting idle, wasting your parallelism.
• One's been looping the whole time and you didn't notice, because it looks busy.
• You alt-tab through a dozen panes reading scrollback to reconstruct who's doing what.

The overhead of tracking the agents starts eating the gains from running them. That's the actual ceiling on parallel agentic coding — not the model, not your machine. Your attention.

Keeping track

A few things help:

• Name everything. Tab titles and worktrees named by task (auth, tests, migration) beat claude — node × 10.
• A tmux dashboard so every session is on one screen instead of behind tabs.
• A per-worktree CLAUDE.md so each agent stays scoped to its task.

But naming and tmux still leave you reading a wall of text. What you actually want is a glance — which agents are alive, which are working, which are stuck, which are done.

That's exactly why we built Mwah: it floats every live Claude Code session on your desktop as its own little robot, showing its real name, project, and live activity — "editing server.ts", "running tests", quiet when idle. One glass for the whole team of agents. You spot the stuck one at a glance, feed the ones that nailed it, and retire the stuck ones with a click instead of hunting for the right terminal.

The playbook

1. Worktree per task, session per worktree — no collisions.
2. Name them by task; keep each one scoped.
3. Watch the board, not the terminals — glance at states instead of reading scrollback.
4. Kill stuck sessions fast — a looping agent is worse than no agent; end it and re-prompt clean. (More: Claude Code stuck in a loop?)

Run agents like a team, not a to-do list. The bottleneck was never how many you could start — it's how many you can see.

---

Related: Shoot your Claude Code in the head · Is your session stuck?