learnaiwithrafa
ClaudeWorkflows

Parallel Claude Agents: Isolation Is Free, Review Is Not

Claude Code now isolates every parallel session in its own git worktree automatically, and Cowork splits big tasks into chunks that run together. Launching agents is solved — here's the setup, plus the review rule that keeps three agents from becoming three messes.

3 min read3 sources
  • #claude-code
  • #parallel-agents
  • #worktrees
  • #cowork

The hard part of running Claude agents in parallel used to be the part nobody screenshots: two agents editing the same folder, silently overwriting each other. That problem is now solved by default — in the Claude Code desktop app, every new session gets its own git worktree automatically, a separate working directory on its own branch. What is not solved, and never will be by tooling, is that every extra agent produces another diff you have to read. Launching is parallel; reviewing is still serial. Plan around that.

The desktop setup, in four moves

  1. Start sessions with Cmd+N (Ctrl+N on Windows). Each session carries its own chat history, context, and — in a git repo — its own worktree, so edits never collide.
  2. Split the screen. Hold Cmd (macOS) or Ctrl (Windows) while clicking a session in the sidebar to open it in a second pane. Cmd+\ closes the focused pane; Ctrl+Tab cycles through sessions.
  3. Make the sidebar readable. Use the controls at the top to group sessions by project, and rename each session (click its title in the toolbar) after the task — not "session 3".
  4. Stop polling. The app fires an OS notification when a session finishes and you aren't viewing it. If you're still tab-flipping every minute, you're paying the old cost for the new feature.

What the worktree isolation actually gives you

Worktrees live under .claude/worktrees/ in your project (add that path to .gitignore). Each one is a fresh checkout on its own branch — the CLI names them worktree-<name> — which means gitignored files like .env don't come along, so the project may not even boot inside a new session. Fix that once: create a .worktreeinclude file in your project root listing the gitignored files to copy into every new worktree. Then enable "Auto-archive after PR merge or close" in Settings so finished sessions clean up after themselves.

Terminal person? claude --worktree feature-auth gives you identical isolation without the app, on a branch named worktree-feature-auth. You can even isolate subagents: add isolation: worktree to a custom subagent's frontmatter and its edits never touch your main checkout.

The same idea outside code: Cowork

Cowork — the Claude desktop tab for non-code work — applies the identical pattern to documents and research: it splits a big task into chunks that run at the same time, researching and organizing while it drafts. It's included on every paid plan, from Pro at $17/month (annual) up through Max, Team, and Enterprise. The test for what to hand it is the same as for code: only tasks that don't need each other's output mid-flight.

The EM angle

Worktrees removed the technical risk of parallel agents, and that's exactly why the human risk grows: nothing stops a dev from launching five sessions and rubber-stamping five diffs at 6pm. The merge step is where quality lives now. On a team, treat agent branches like junior-engineer PRs — same review bar, and no exception because "the agent wrote it". Two agents whose work you actually verified beat five you skimmed.

Today: update Claude Desktop, add .claude/worktrees/ to your .gitignore, create a .worktreeinclude with your .env, and run exactly two parallel sessions on tasks that don't touch — one real feature, one chore you can verify in minutes.

Sources