Parallel agent work runs into the file system, not the model. Two agents in one working directory inevitably step on each other's edits, and no amount of careful wording cures that: they read a file, think about it for a few seconds and write the result back, unaware that the file changed in the meantime. A git worktree solves the problem physically: a separate checkout with its own files, its own branch and its own dependencies. That is why parallelism starts with trees rather than with the number of open chats.

The naive scheme is to launch two agents in one folder and hope they will divide the files. It breaks even without git conflicts: one changes a contract while the other simultaneously adapts to the old version, and the result is a semantic conflict the version control system will not see. Formally everything merged, in fact the behavior is broken - and untangling that costs more than separating the tasks into trees in the first place. Git compares lines, not meanings, and where two edits did not overlap textually it will honestly report success.

Cursor provides trees both through the agents window and through a set of commands in the editor: create a tree for a task, apply the result, delete the tree. A separate technique is to run the same task on several models, where each candidate works in its own tree. That is an honest way to compare approaches without mixing them: the results are visible side by side rather than layered on each other in a shared directory. The value is not that one model turns out to be better in general, but that on this particular task several solutions are visible at once and there is something to choose between.

An important detail about comparing candidates: the winner does not merge itself. After choosing, you have to review the changes and commit or apply the tree by hand. That is not a shortcoming but a deliberate boundary - choosing among several variants remains a human decision rather than an automatic action based on a comparison. Automation could not decide correctly here anyway: the criterion usually lies outside the code, in which approach you are prepared to maintain from now on.

It helps to see the setup of a new tree once. Below is a configuration with preparation commands for different platforms. Cursor reads it first from the tree itself, then from the project root; the value can be either an array of commands or a path to a script next to the configuration. The reading order matters: it lets a tree created for a particular task prepare itself differently from the project default without touching the shared configuration.