A parent or planner agent breaks a task into independent pieces and dispatches each to a subagent running on its own isolated VM or sandbox, so parallel edits and test runs cannot collide on shared state. This is closer to fan-out task parallelism with a coordinator than to autonomous multi-agent negotiation — the planner still decides how work is split, and the swarm works best on independent, low-coordination slices of work rather than tasks with sequential dependencies.