git worktree Is the Multi-Agent Parallelism Primitive AI Coding Demands
AI coding agents are inherently concurrent — you want one fixing a bug while another builds a feature — but Git's default single-workdir model forces serial execution and risks cross-contamination. worktree provides the filesystem isolation that makes parallel agent workflows safe and practical, without changing how Git itself operates.
The single-working-directory model breaks down hard when AI agents enter a codebase: one agent switching branches can swallow your uncommitted work, and two agents running simultaneously overwrite each other's files. git worktree solves this by attaching separate directories to the same .git object database, one per branch, so each agent gets an isolated filesystem while sharing commit history. The main working tree stays clean and releasable no matter what intermediate junk the agents produce, and your IDE stops getting bombarded with external file-change notifications.
The workflow is straightforward: `git worktree add -b agent/task ../task-dir` spawns a new directory checked out to a unique branch. Agents run confined to that directory; you review with `git diff`, merge back from the main tree, and remove the worktree when done. The constraint that a branch can only be checked out in one worktree at a time enforces a clean one-task-one-branch-one-directory mapping.
Pitfalls include dependency directories (node_modules, .venv) not following the worktree, agents accidentally wandering into sibling directories via relative paths, and the need to manually diff AI output before merging. But the core payoff is immediate: multiple agents run concurrently, the main branch never gets polluted, and context-switching no longer requires stash stacking.
git worktree has existed since 2015 but remained a niche parallel-feature trick until AI coding agents made concurrent, isolated working directories a daily necessity rather than an occasional convenience.
The constraint that a branch cannot be attached to two worktrees simultaneously is not a limitation in practice — it forces a clean mapping that prevents the exact confusion (wrong branch commits, lost context) that stash-based workflows invite.
The main tree's persistent cleanliness under worktree isolation changes the developer's relationship with CI and releases: the question "can I ship right now?" becomes answerable at a glance, regardless of how many agents are running.
Agent isolation via worktree shifts the review burden from "recover what the agent broke in my working directory" to a deliberate diff-and-merge step, keeping the human in the decision loop without slowing parallel execution.