Worktrunk makes git worktrees practical enough to run 10 AI agents at once

Git worktrees have always been the right answer for running multiple AI agents in parallel — each agent gets its own working directory, no collisions, no stashed state. The problem is that native git worktree commands are genuinely painful to type. Creating a single worktree requires specifying the branch name three times, the path separately, then changing into the new directory. Worktrunk collapses that into a single command: wt switch --create feature-auth creates the branch, creates the worktree at a computed path, and drops you there.

The tool is built in Rust and positions itself squarely around the agentic coding workflow. The three-command demo in the README — switch, list, remove — covers the lifecycle that matters when you're spinning up 5 to 10 parallel Claude or Codex sessions. The -x flag is the key piece: wt switch -x claude -c feature-a -- 'Add user authentication' creates the worktree and immediately launches the agent inside it, passing the task description as an argument. That's the whole loop in one line.

wt list is more useful than it first appears. The output shows staged changes, commits ahead or behind main, unpushed commits, CI status, and optionally LLM-generated summaries per branch — all in a single table. When you have eight worktrees open, knowing which ones are blocked on CI and which are ready to merge without switching into each one is genuinely valuable.

The merge workflow is worth noting separately. wt merge main generates a commit message from the diff, squashes if needed, rebases onto main, fast-forward merges, and removes the worktree and branch — all in sequence, with output at each step. That's the kind of automation that makes the parallel-agent pattern actually sustainable rather than a novelty that collapses under its own bookkeeping.

The hook system handles the setup friction that worktrees introduce. Every new worktree starts cold: no node_modules, no compiled target/ directory. Worktrunk's copy-ignored-files step can clone those directories from an existing worktree, and post-create hooks can run install commands automatically. Without something like this, the cold-start cost per agent session quietly eats the time you saved by parallelizing.

The Windows install has a real naming conflict with Windows Terminal's built-in wt command, which the Winget package sidesteps by installing as git-wt instead — a friction the README addresses honestly. The tool describes itself as the most popular git worktree manager; that claim is unverifiable from the README alone, but Anthropic's own best-practices guide links to the worktree pattern, and incident.io has published a workflow built around it.