Worktree

A worktree is an extra checkout of a git repository in a separate directory, sharing the same .git history but sitting on its own branch. git worktree add ../feature-x feature-x gives you a second working copy where a change can be made, built and tested without disturbing the files in your main checkout. For an agent, a worktree is a private filesystem to work in while you keep using yours.

The problem it solves shows up the first time you run two things at once. You start an agent on a refactor, then open the same repo to fix a bug; the agent's half-finished edits are in your editor, your fix lands in its context mid-task, and both of you end up confused about which files are whose. Branch switching doesn't help, because a switch rewrites the very files the agent is editing. Separate directories do.

Give each parallel task its own worktree, especially for AFK runs, and treat a worktree like a sandbox with the code already inside. Remember they cost disk space and a fresh dependency install each, so prune finished ones (git worktree remove) and commit or discard the work before you do. Some harnesses create and clean up worktrees for you per task; that is the same mechanism with the housekeeping automated.