How to run multiple AI coding agents in parallel with git worktrees

One of the biggest shifts in AI-assisted development through 2026 is that you no longer point a single agent at a single task. Tools like Claude Code and Cursor now encourage running several agents in parallel — one fixing a bug, another writing tests, a third drafting a feature. The catch: if they all edit the same working directory, they overwrite each other's changes and corrupt the build.

The clean fix is an old Git feature that turns out to be the perfect primitive for this: worktrees.

What a worktree actually is

Normally a Git repository has one working directory — the folder you edit files in. git worktree lets one repository have multiple working directories at once, each checked out to a different branch, all sharing the same underlying .git history. Each worktree is a real folder on disk with its own files, so an agent working in one can't touch the files in another.

Think of it as giving every agent its own desk in the same office, instead of three people fighting over one keyboard.

The core workflow

From inside your repo, create a worktree for each task on its own branch:

# Create a new branch + folder for an auth bug fix
git worktree add ../myapp-auth-fix -b fix/auth-bug

# Another for a feature, in parallel
git worktree add ../myapp-search -b feat/search

# See everything you have checked out
git worktree list

You now have two sibling folders next to your main repo: myapp-auth-fix and myapp-search. Open a separate agent session in each — cd ../myapp-auth-fix and start Claude Code there, then do the same in ../myapp-search in another terminal tab. Each agent has a full, independent checkout. They run at the same time, commit to their own branches, and never collide.

When a task is done, merge its branch and remove the worktree to keep things tidy:

git checkout main
git merge fix/auth-bug
git worktree remove ../myapp-auth-fix

A concrete example

Say you're building a small web app and you have three independent jobs: fix a flaky login redirect, add pagination to a list, and bump a slow dependency. Instead of doing them one at a time:

git worktree add ../app-login   -b fix/login-redirect
git worktree add ../app-paging  -b feat/pagination
git worktree add ../app-deps    -b chore/upgrade-deps

Launch an agent in each folder with a focused prompt. While the login agent traces the redirect logic, the pagination agent is already writing components, and the dependency agent is running the upgrade and the test suite. Twenty minutes later you review three separate, clean branches and merge the ones that pass.

Tips that keep it sane

• Keep tasks independent. Worktrees prevent file conflicts, but if two agents edit the same files on different branches, you'll still hit merge conflicts at the end. Split work along natural seams.
• Give each worktree the same context. A CLAUDE.md (or equivalent rules file) committed to your repo travels into every worktree automatically, so each agent inherits your conventions.
• Watch the dependency folders. node_modules, virtual environments, and build caches are not shared between worktrees. You may need to run npm install once per worktree, or use a package manager with a global store (pnpm, uv) to avoid re-downloading everything.
• Don't over-parallelize. In practice, three to five concurrent agents is the sweet spot. Beyond that, the time you spend reviewing and coordinating outweighs the speed you gain.

Native worktree integrations have been landing across the major tools since late 2025, and dedicated launchers now exist to spin sessions up for you. But you don't need any of them to start — git worktree add is built into Git, works today, and is the single most reliable way to let several AI agents build for you at once without tripping over each other.