Three Agents in Parallel on Git Worktrees — Agent Engineering

Step 1 of 4

Lab brief

Goal

Three merged PRs from parallel agents with no file collisions, plus a note on where review became the bottleneck.

Failure mode

Three agents in one checkout overwrite each other's edits, fight over the dev-server port, and produce a tangled diff nobody can review.

Staff move

Parallelize only independent, specified tasks; isolate each in its own worktree and port; keep PRs small so review keeps up.

Step 2 of 4

1. Split the work

Pick a milestone with three tasks that touch different files (for example: a new API endpoint, a UI badge, docs)
Write a 5-line spec per task with acceptance criteria
Check for shared files; if two tasks touch the same module, run them sequentially instead

Step 3 of 4

2. Create worktrees and launch

Setup

git worktree add ../app-api    -b feat/api-endpoint
git worktree add ../app-badge  -b feat/ui-badge
git worktree add ../app-docs   -b docs/endpoint

# one terminal (or editor window) per worktree
cd ../app-api   && PORT=3001 <start your agent with spec 1>
cd ../app-badge && PORT=3002 <start your agent with spec 2>
cd ../app-docs  &&            <start your agent with spec 3>

Cursor and the Codex app can manage worktrees for parallel agents for you; cloud agents are another option when the tasks do not need your local environment.

Step 4 of 4

3. Review and measure

Review each PR with a clean-context reviewer first, then yourself
Record wall-clock time vs your estimate for doing the three serially
Clean up: git worktree remove ../app-docs after merge

Stretch

Best-of-N: run the hardest task in two extra worktrees with different models and keep the best diff.