Running two AI coding agents on the same repository? Then they're probably mixing up each other's changes. The fix is a git feature that has existed since Git 2.5 (2015) and that few people used until AI agents arrived: git worktree.
Everything below was tested with git 2.50, on a small demo repo and on the real express repository.
Video version:
The problem: one folder = one working directory
Agent A is halfway through a login fix in index.js. Agent B starts its own task in the same folder:
# agent A has an uncommitted change in index.js…
git checkout -b agentB
echo "// B: search" >> index.js
git commit -am "B: search"
git show HEAD -- index.js
module.exports = require('./lib/express');
+// A: login fix ← agent A's unfinished work!
+// B: search
Switching branches keeps uncommitted changes, so agent B's commit now contains agent A's half-done work. It gets worse when both agents edit the same file at the same moment, or when one runs git stash or git checkout . and wipes out the other's changes.
The fix: git worktree add
git worktree add ../shop-login -b feat/login
git worktree add ../shop-search -b feat/search
Each command creates a new folder with its own files and its own branch, connected to the same repository:
$ git worktree list
~/code/shop 71d6cb4 [main]
~/code/shop-login 71d6cb4 [feat/login]
~/code/shop-search 71d6cb4 [feat/search]
Open one agent in each folder. They can edit, run tests and commit at the same time without touching each other's files.
cd ../shop-login && claude # or Codex, Copilot CLI, Aider…
Inside a worktree, .git is just a small file pointing back to the main repository:
$ cat ../shop-login/.git
gitdir: ~/code/shop/.git/worktrees/shop-login
Why not just clone the repo again?
Because worktrees share the repository: history, branches and objects.
On express (6,186 commits):
| Disk | |
|---|---|
Second git clone
|
13 MB |
git worktree add |
1.3 MB |
And because it's one repository, a commit made in one worktree is instantly visible from all the others. I committed in the agent's worktree, and git log agent1 in the main folder showed it right away, with no push or pull between copies.
Git's built-in safety checks
A branch can only be checked out in one worktree at a time:
$ git checkout feat/login
fatal: 'feat/login' is already used by worktree at '~/code/shop-login'
And git won't delete a worktree with unsaved work:
$ git worktree remove ../shop-login
fatal: '../shop-login' contains modified or untracked files, use --force to delete it
Gotchas
-
Untracked files aren't copied. A new worktree has only tracked files, so
.env,node_modulesand build output are missing. Runnpm installand copy your.envinto each worktree. - Ports: if each agent starts a dev server, give each one a different port.
- Merge conflicts don't disappear. If two tasks change the same files, you'll resolve conflicts when you merge, just later.
- Review still matters. Worktrees isolate the work; they don't check it. Review each branch before merging.
Cheat sheet
| Command | What it does |
|---|---|
git worktree add ../dir -b branch |
New folder + new branch |
git worktree add ../dir branch |
New folder for an existing branch |
git worktree list |
Show all worktrees |
git worktree remove ../dir |
Delete a worktree (after merging) |
git worktree prune |
Clean up worktrees whose folders you deleted by hand |
Some AI tools can create worktrees for you. For example, Claude Code has claude --worktree [name], which starts a session in a fresh worktree.
How many AI agents do you run at once? 👇
I post one tested coding tip every day on YouTube at Refactorly. The 30-second version of this one is here.
Top comments (0)