Taufiq Septryana
dev kodi

Git Worktree: The Key to Parallel AI Agent Development

Learned this the hard way today: when spawning multiple AI sub-agents to work on different branches of the same repo simultaneously, they all share the same working directory. This means they stomp on each other’s changes — one agent checks out branch A, another checks out branch B, and suddenly both are confused.

The fix: git worktree.

The Problem

I spawned three sub-agents to create PRs for three different features (#38, #39, #40) in the same repo. They all worked in the same clone directory, leading to branch conflicts and tangled changes that needed painful manual untangling.

The Solution

# Create isolated working directories per branch
git worktree add ../repo-feature-a feat/feature-a
git worktree add ../repo-feature-b feat/feature-b
git worktree add ../repo-feature-c feat/feature-c

Each worktree gets its own directory with its own checked-out branch, but they all share the same .git history. No conflicts, no stomping — true parallel development.

Why This Matters for AI Agents

This pattern is especially relevant for AI coding agents working in parallel. Whether it’s Cursor, Claude Code, or custom sub-agents:

  • Each agent gets an isolated workspace
  • No git checkout conflicts between agents
  • All branches stay in sync through the shared git object store
  • Cleanup is simple: git worktree remove ../repo-feature-a

The takeaway: Before spawning parallel agents on the same repo, always set up worktrees first. Five seconds of setup saves thirty minutes of untangling.

Back to TIL