git worktree で AI マルチタスクを隔離する — チャット分けだけでは足りない理由Using separate Cursor chats is not enough for true AI multitasking isolation,…

Running two or more AI coding sessions at once has become a common way to squeeze more out of tools like Cursor, but the practice exposes a subtle problem that many developers only notice when it is too late. Opening a second chat to handle a separate task feels like opening a second workspace. In reality, both conversations are almost always operating on the same Git working tree, which means they share a single branch state. That shared state is where parallel work quietly collides.

A concrete example illustrates the trap. Imagine one chat is fixing authentication logic while another drafts a blog article. On screen, the two tasks appear to progress independently and calmly. Then, just before committing, a quick git status reveals that untracked Markdown from the writing task is sitting right next to the authentication diff. Worse, the branch that was supposed to hold the auth work has, at some point, switched over to the blog branch. Nothing warned you; the working tree simply reflected whichever branch was checked out last.

The underlying reason is straightforward once stated. A repository has one working tree, and a working tree can have only one branch checked out at a time. Each chat may believe it is operating on its own branch, but a git switch issued on behalf of one task changes the branch for everyone looking at that directory. The other task does not get its own copy of the files; it inherits whatever the last switch produced. Uncommitted changes, staged files, and untracked artifacts all pool together in the same physical location, so isolating tasks by simply opening separate conversations is not enough for true multitasking isolation.

This is the gap that git worktree is designed to close. The worktree feature, part of Git for several years now, lets a single repository check out multiple branches into multiple directories simultaneously. Instead of one working tree bound to one branch, you create additional linked working trees, each with its own directory on disk and its own checked-out branch. A command such as git worktree add ../feature-auth feature/auth produces a new folder that tracks the authentication branch, while the original directory keeps working on something else. The two directories share the same underlying object database and history, so no repository is duplicated, but the files, the staging area, and the current branch are physically separated.

Applied to the AI workflow, the fix is to give each task its own directory and point each chat at a different worktree. The authentication session lives in one folder, the blog draft in another. A git switch in one no longer disturbs the other, and git status in either directory shows only the changes that belong to that task. Untracked drafts can no longer wander into an unrelated commit, because they never occupy the same working tree in the first place. The isolation is enforced by the filesystem rather than by the developer's memory.

There are a few practical points worth keeping in mind. Git prevents the same branch from being checked out in two worktrees at once, which is a safeguard rather than a limitation. Worktrees do consume additional disk space for their working files, and they need to be cleaned up with git worktree remove or pruned when a task is finished, otherwise stale directories accumulate. Editors and AI tools also need to be opened against the correct folder so that their context matches the intended branch; the separation only holds if each session is genuinely rooted in its own worktree.

It is also useful to see where this sits among adjacent approaches. Some teams reach for separate clones of a repository to achieve isolation, but full clones duplicate history and are heavier to maintain, whereas worktrees share one object store. Container-based or devcontainer setups can isolate environments and dependencies, which is a related but distinct concern from branch and working-tree state. As agentic and multi-session AI coding grows more common, the working tree increasingly appears to be the real bottleneck for parallelism, and git worktree offers a low-cost, native way to give each concurrent task the clean, independent surface it needs.