Run agents in parallel with git worktrees

Two agents in one repo mostly fail by stepping on each other's working tree. A worktree gives each task its own directory and branch.

git worktree add ../repo-task-a -b task-a opens a second working directory backed by the same .git. Each agent works in its own; edits, test runs and installs stop overwriting each other.

One task, one worktree, one branch. Do not let two agents share a dirty tree or edit the same file. The mess appears before you read a diff, not at merge time.

node_modules, .env and build caches do not follow a new worktree. Reinstall and copy local config there, and write that down in the repo brief so the agent does not guess in an empty directory.

When the task is done, git worktree remove, then delete the branch. Stale worktrees are where an agent edits six-month-old code. Cap parallel tasks at the number of diffs you can actually read.