Run five agents at once. None of them stepping on each other.

Two agents in the same working directory will eventually collide. Agent A edits a file agent B is reading. Agent B commits something that conflicts with agent A's in-progress changes. You end up with a merge conflict in the middle of an automated run, both agents stopped, and a working directory in an unknown state.

The common workaround is to run agents sequentially. One finishes, the next starts. But sequential execution erases the entire point. You have turned a parallelism tool into a slower version of one agent.

Git worktrees are separate checked-out working directories linked to the same repository. Each worktree has its own branch and its own file state. Changes in one do not affect any other until you explicitly merge.

This is the correct way to isolate parallel agents. Each agent lives in its own worktree, works on its own branch, and cannot touch anything outside its directory. The blast radius of a bad agent run is zero: it cannot corrupt main, collide with another agent, or leave your working state broken.

Tempest creates and manages these worktrees automatically. Open a session tab — worktree created. Close it — worktree persists. You never touch the Git worktree CLI directly.

When you run multiple agents in parallel, you are not watching all of them at once. You delegate, do something else, then come back to review. If coming back costs five minutes of re-orientation per session, five sessions costs twenty-five minutes before you have reviewed a single diff.

Tempest persists every session the moment you leave it. Close a tab, reopen it — the agent picks up exactly where it left off. Full conversation history, same branch, worktree untouched. Nothing is lost because nothing was lost.