split a job across several agents
The work can run at the same time, not as one long thread. You need a model that can own a slice and a model that can merge the slices when they disagree.
recs
| model | best when | avoid when | cost vibe |
|---|---|---|---|
| Frontier | The slices touch the same files, or someone has to decide which result wins. | Each slice is independent and the merge is a checklist. | $$$ |
| Faster / cheaper cloud | The cuts don't overlap, and a failed slice can be thrown out without touching the others. | Two agents will edit the same function and you have no owner for the conflict. | $ |
| Local mid | The slices are small, private, and you are the merge. | The fan-out is wide enough that you're just retyping the same prompt five times. | ~free |
why
- Parallel fails by collision, not by a slow first answer. Two agents in the same file is the bug.
- Cheap models are fine on slices that don't meet. Frontier earns it on the merge, when the pieces disagree.
- Name the owner of each file before you fan out. A job with one shared file is not parallel.
- Use long-autonomous-task when the work is sequential. Use this page when the slices can actually run together.