How do I run parallel AI coding agents with git worktrees?

Give every agent its own git worktree: a separate checkout on its own branch that shares the repository's history. With plain git, run git worktree add ../app-feature-a -b feature-a and start the agent in that folder. With Claude Code, run claude --worktree feature-a, which creates .claude/worktrees/feature-a on a new branch worktree-feature-a; repeat with another name in a second terminal. Review each branch, merge the ones that work, then remove the worktrees.

Why use git worktrees instead of several clones?

Worktrees share one .git directory, so they are created in seconds, use less disk, and every branch an agent makes is immediately visible from your main checkout. Separate clones work too, but you have to fetch between them. Either way, the point is that two agents never edit the same files on disk.

How do I get .env files into a new worktree?

A worktree is a fresh checkout, so gitignored files such as .env are missing. Claude Code copies files listed in a .worktreeinclude file into each new worktree. With plain git, copy them in a setup script, or use a tool such as Worktrunk or a desktop orchestrator that runs a setup hook for each workspace.

How many coding agents can I run at once?

The limit is usually your review time and your plan's usage, not the machine. Three to five agents on independent tasks is a common working range on one laptop; beyond that you spend your day reviewing diffs. Each agent also needs its own dev-server port and, if tests touch a database, its own database or schema.

Quick Answer

How to Run Parallel AI Coding Agents With Git Worktrees

Published: November 6, 2026

The short answer

One task, one agent, one worktree. A git worktree is an extra working directory on its own branch that shares your repository’s history. Put each agent in its own worktree and they cannot overwrite each other’s files. As of November 2026 the fastest way is:

# Claude Code: one isolated session per terminal
claude --worktree feature-auth
claude --worktree fix-billing-bug

# Any agent, plain git
git worktree add ../myapp-feature-auth -b feature-auth
cd ../myapp-feature-auth && codex

Then review each branch, merge what works and delete the rest.

Step 1: Prepare the repository

Commit first. Worktrees are created from a commit; in a repository with no commits, claude --worktree fails with Failed to resolve base branch "HEAD".
Ignore the worktree folder. Claude Code creates worktrees under .claude/worktrees/<name>/; add .claude/worktrees/ to .gitignore so they do not show up as untracked files.
Split the work. Agents in parallel only help on tasks that touch different files: a feature, a bug fix, a test backfill, a dependency bump. Two agents refactoring the same module will produce a merge conflict you have to resolve yourself.

Step 2: Create one worktree per agent

Claude Code. claude --worktree <name> (or -w) creates .claude/worktrees/<name>/ on a new branch worktree-<name> and starts Claude there. Omit the name and Claude invents one. Inside a session you can also say “work in a worktree” and Claude creates one with its EnterWorktree tool. Interactive runs need workspace trust, so run claude once in the repository first.

Plain git, for any agent (Codex, Gemini CLI, OpenCode, Aider…):

git worktree add ../myapp-task-1 -b task-1
git worktree add ../myapp-task-2 -b task-2
git worktree list

Worktrunk makes this one command and starts the agent too: wt switch -c -x claude feature-a creates the branch and worktree and launches Claude in it (our Worktrunk review).

Step 3: Set up each worktree’s environment

A worktree is a fresh checkout. Before the agent runs tests it needs:

Gitignored files — .env, local certificates. Claude Code copies anything listed in .worktreeinclude into every new worktree. With plain git, a two-line setup script does it.
Dependencies — npm install, uv sync or equivalent, per worktree. Ask the agent to do it as its first step.
Separate ports — two dev servers cannot both bind port 3000. Pass PORT=3001 and so on.
Separate test state — if tests write to a local database, give each worktree its own database name or schema.

Step 4: Watch and review

Keep one terminal per agent, or use something that shows them side by side. Claude Code’s background agents let you monitor parallel sessions from one screen. Desktop apps such as Orca, Paradev and T3 Code show each agent’s status and diff and open the PR for you; they replace Conductor, whose cloud service shuts down on November 20, 2026 after its team joined Meta (Conductor alternatives).

Review each branch like a colleague’s PR: run the tests yourself, read the diff, and throw away branches that went sideways. Rerunning a task in a fresh worktree is usually cheaper than rescuing a confused one.

Step 5: Merge and clean up

git switch main
git merge --no-ff task-1        # or open a PR from the branch
git worktree remove ../myapp-task-1
git branch -d task-1
git worktree prune               # drop entries for folders you deleted by hand

When you exit a Claude Code worktree session, Claude checks it for uncommitted files and new commits: a clean worktree from an unnamed session is removed automatically with its branch; otherwise Claude asks whether to keep or remove it, and prints a claude --worktree <name> --resume command if you keep it. With Worktrunk, wt merge main and wt remove do the same.

Common problems

Problem	Fix
”fatal: ‘feature-a’ is already checked out”	A branch can be checked out in only one worktree. Use a new branch name per agent.
Agent edits the main checkout instead	Start the agent from inside the worktree folder; for Claude Code use `--worktree` rather than `cd`.
Tests fail only in the worktree	Missing `.env` or dependencies: add them to `.worktreeinclude` or your setup script.
Disk filling up	Each worktree holds its own `node_modules` or virtualenv; remove finished worktrees promptly.
Two agents changed the same file	Merge one branch, rebase the other on `main`, and let the second agent resolve the conflict.

When worktrees are not enough

Worktrees isolate files on one machine. Long tasks that should run while your laptop sleeps belong in cloud agents (Codex cloud, Cursor cloud agents, Claude Code cloud sessions); see best AI coding agents for parallel development. And parallel agents use your plan’s allowance in parallel, so watch usage before scaling from three agents to ten.

Last verified: November 6, 2026, against the Claude Code worktree documentation, the git-worktree manual and the Worktrunk README.