How Claude Code’s Native Git Worktree Support Enables Fully Parallel Multi‑Agent Workflows

On February 21, 2026, Boris Cherny, head of Claude Code, announced that the CLI now ships with built‑in Git Worktree support.

Problem: When several AI agents run concurrently on different tasks (e.g., refactoring a database module and fixing a login bug) they share the same repository, branch, and HEAD. One agent may modify utils.py while another reads it, leading to version clashes, index contamination, and even data loss.

Git Worktree is a long‑standing Git feature that allows a single repository to host multiple independent working directories, each with its own branch, HEAD, and index while sharing the same .git object store.

"A single company (the repository) opens multiple independent offices (worktrees). Each office has its own whiteboard and progress board, but everyone shares the same archive (.git) where file history is stored."

Typical usage:

# Create a new worktree with a new branch
git worktree add ../my-feature-branch -b feature-login

# Remove the worktree after work is done
git worktree remove ../my-feature-branch

Claude Code integrates this capability directly into its CLI. By passing the --worktree flag, the tool creates an isolated worktree for the session: claude --worktree my_worktree If no name is supplied, Claude auto‑generates one and creates the directory under .claude/worktrees/. The same isolation can be enabled in the GUI by checking the “worktree mode” option on the Code tab.

Subagents now also inherit worktree isolation, allowing heavy users to launch many agents in parallel without manual setup. Boris illustrated this with a scenario where ten agents each run in their own worktree, produce ten independent PRs, and modify disjoint parts of the codebase (e.g., src/utils/ vs. vendor/) without any interference.

Custom agents can declare worktree isolation in their front‑matter:

---
name: worktree‑worker
model: haiku
isolation: worktree
---

The update is not limited to Git. By configuring hooks in .claude/settings.json, similar isolation can be achieved for Mercurial, Perforce, SVN, or any VCS that can execute a command when a worktree is created or removed.

{
  "hooks": {
    "WorktreeCreate": [{"command": "jj workspace add \"$(cat /dev/stdin | jq -r '.name')\""}],
    "WorktreeRemove": [{"command": "jj workspace forget \"$(cat /dev/stdin | jq -r '.worktree_path')\""}]
  }
}

On the same day, Vue.js creator Yuxi Yu posted a “cyber cricket” test on X, running Claude Code 4.6 Opus and Codex 5.3 xhigh in two separate worktrees on identical tasks. The result showed Opus outperforming Codex, mainly because Opus pinpointed an upstream bug in a dependency while Codex simply bypassed the failure.

The test is informal and not strictly controlled, but it demonstrates that parallel worktree workflows make side‑by‑side AI‑coding tool comparisons easy and natural.

From a practical standpoint, the update brings two major benefits: (1) it dramatically reduces friction when running multiple agents in parallel, eliminating the need to manually clone repositories or coordinate branches; (2) it provides a stable infrastructure for large‑scale code migrations, bulk refactoring, and concurrent experimentation. However, worktree isolation only solves workspace conflicts; the quality of each agent’s output and the merging of results still require human judgment.