Explained

Git worktrees for AI coding agents

A git worktree is a second working copy of the same repository, on a branch of its own, so two agents can edit at once without touching each other's files. It keeps them apart while they work and moves the meeting point to the merge, where their changes have to be put together.

Setting one up

git worktree add ../app-codex -b codex/checkout-flow
cd ../app-codex && npm install
codex   # or claude, in its own terminal

Each worktree is a folder of its own. It shares the repository’s history with the others but not its files, so it needs its own dependency install, its own environment file, and its own ports if two of them run a development server.

What it costs

Worktrees keep agents from writing the same file. They also keep them from seeing each other’s work: the agent building the interface cannot call the endpoint the other one wrote until that branch is merged. Two agents that both had to touch a shared file, such as the types or the routes, each made their own version of it, and the merge is where those meet.

The work around it is splitting tasks so they rarely touch the same files, merging small and often, and giving shared files a single owner.

The other way: one directory and a reservation per file

Murmell keeps every agent in one shared directory on a cloud machine. An agent reserves the file it is about to write, and the others wait for it rather than writing over it. They see each other’s code as soon as it is saved, and there is no branch per agent to merge. Multiple AI agents on one repo has the mechanism, and tools to run coding agents in parallel compares the tools built on each approach.

Straight answers