Parallel Agents in Isolated Workspaces
AssessFactorial rates Parallel agents as Assess (2026). Some of us run two or three agents in parallel, each in its own jj workspace or git worktree, and use a terminal multiplexer to keep track of them.
Coding agents spend much of their time waiting for builds, tests and model responses. Running several at once, each on its own task, makes better use of that time. To keep them from editing the same files, every agent gets its own checkout of the repository: a git worktree or a Jujutsu workspace, with a separate working copy on the same history.
The checkouts are cheap to create and throw away, and every agent's work ends up as an ordinary change or branch that can be reviewed and merged on its own. What does not get cheaper is the merge itself and the human attention needed to review several streams of work.
Updates
Some of us run two or three agents in parallel, each in its own jj workspace or git worktree, and use a terminal multiplexer to keep track of them. Our skills create and remove the workspaces, including per-app checkouts in our multi-repo client workspaces. It works well for independent tasks such as a bug fix next to a feature. It does not work well when tasks touch the same code: the merge conflicts and the review load land on the developer, and that review is where the time goes (see human review of agent diffs).