How to Run Multiple Claude Instances in Parallel | Agent Teams, tmux, and Worktree

What Is "Multi Claude"?

"Multi Claude" is a collective term for techniques that run multiple Claude instances simultaneously to process tasks in parallel. The official Claude Code documentation (source: https://code.claude.com/docs/ja/agents) classifies this parallelization into four approaches. The choice between them comes down to three axes: who coordinates the work, whether workers communicate with each other, and whether they touch the same files.

In the past, the dominant style was coding through a single conversational session, but as Claude Code has matured, there are increasingly more cases of running multiple processes simultaneously to multiply raw development speed. The key caveat is that token usage increases linearly as the number of sessions grows.

Comparing the Four Approaches

Here is a summary of the four approaches described in the official documentation.

Additionally, there are two supporting tools: Worktree (provides an independent git checkout for each parallel session) and the /batch command (splits a single change across 5–30 Worktree-isolated sub-agents).

Sub-Agents — Delegated Execution Within a Single Session

This is the easiest parallelization method. The main Claude session delegates a specific side task to a separate Claude instance, which processes it in its own context window. The result is returned as a summary to the main session, so the main conversation never gets bloated with unnecessary information.

The command to use is /agents, which opens a panel showing currently running sub-agents along with a Library tab for creating and editing custom sub-agents. By placing Markdown files with YAML frontmatter in .claude/agents/, you can define custom sub-agents that can be shared with your team.

Sub-agents represent division of labor within a single session — they do not spin up completely independent separate processes. State between them is shared only through the main session.

Agent View — Centralized Monitoring of Background Sessions

Agent View, launched with the claude agents command, is a dedicated screen for dispatching and monitoring sessions running in the background (currently in research preview).

Here is how it works:

1. Run claude agents to open the Agent View screen.
2. Dispatch a new background session.
3. Check the status of each session (running, waiting, awaiting input) at a glance.
4. Attach to and intervene in only the sessions that need input.

Each session dispatched by Agent View automatically moves to its own Worktree, so you can run parallel processing without worrying about file conflicts. It is ideal for "running multiple independent research tasks in parallel and reviewing them all at once when you have time."

Agent Teams — A Leader Driving a Team (Experimental)

Agent Teams is a system in which a leader session plans, assigns, and supervises multiple teammate sessions. At this time, it is an experimental feature and is disabled by default.

Teammates share a task list and can send messages directly between agents. However, Agent Teams does not automatically isolate teammates in Worktrees during testing, so you need to design your task breakdown carefully to ensure multiple teammates do not touch the same set of files (source: https://start-link.jp/hubspot-ai/ai/claude-code-practice/claude-code-parallel-agents).

While well-suited for large-scale project splitting, note that its experimental status means the behavior specifications may change.

Running Independent Sessions in Parallel with tmux and Multiple Terminals

If you want to "handle tasks from different repositories at the same time" or "run completely independent processes," splitting terminals with tmux is the simplest solution.

The basic steps are as follows:

1. Create a new session with tmux new-session.
2. Add panes with Ctrl+b % (vertical split) or Ctrl+b " (horizontal split).
3. Launch the claude command in each pane and assign separate tasks.
4. Move between panes with Ctrl+b o to check progress.

With tmux, each Claude session runs as a completely independent process. While one session is handling a heavy task, the other can continue a separate conversation. When working in the same repository, combining tmux with Worktree prevents file conflicts.

Preventing File Conflicts with Worktree

When multiple Claude sessions edit the same files in the same repository, conflicts and unexpected overwrites can occur. Git worktree is the fundamental solution to this problem.

The setup steps are as follows:

1. Add a Worktree with git worktree add ../project-task-b feature/task-b.
2. Launch a separate Claude session in the created ../project-task-b directory.
3. The main working tree and ../project-task-b share the same git repository while each maintains an independent file tree.
4. After finishing, remove it with git worktree remove ../project-task-b.

Since Agent View automatically moves dispatched sessions to a Worktree, you do not need to do this manually. If you are self-managing with tmux, it is safer to set up Worktree manually.

Splitting Large-Scale Changes in Parallel with the /batch Command

/batch is a Claude Code skill that automatically splits a large change into 5–30 Worktree-isolated sub-agents. Each sub-agent processes its work in an independent Worktree and creates a pull request upon completion.

For cases like "refactoring a specific pattern across the entire codebase" or "applying the same migration to a large number of files," this can dramatically reduce time compared to manual or sequential processing. Note that this is a packaged use of sub-agents and Worktree — it is not a distinct coordination style.

Managing Costs for Parallel Execution

Running multiple Claude sessions simultaneously causes token usage to increase proportionally with the number of sessions. Expect roughly three times the consumption when running three sessions at once.

The basic principles for keeping costs down are:

1. Start with 1–2 parallel sessions and measure the effect before scaling.
2. Do not apply parallelization to tasks that are not fully independent (i.e., those with dependencies).
3. Be aware that the experimental Agent Teams feature can inadvertently generate many sub-sessions.
4. For large-scale processing (/batch), confirm the scope before running.

You can check usage details with the /costs command in Claude Code or on the costs page of the official documentation (source: https://code.claude.com/docs/ja/agents).

Summary — Choosing by Use Case

The four "multi Claude" approaches are easiest to differentiate by task independence and scale.

• Don't want side tasks mixed into the main conversation → Sub-agents (/agents)
• Want to hand off multiple tasks and review them later → Agent View (claude agents)
• Want Claude to form a team structure and manage a project → Agent Teams (experimental)
• Want to process a 500-file migration all at once → Dynamic workflows or /batch
• Want to run independent processes across different repositories → tmux + multiple sessions

Worktree serves as "file conflict prevention" regardless of which approach it is combined with. Except in cases where Claude Code automates this (Agent View), it is safest to set up Worktree manually whenever you run multiple sessions in the same repository.