ultrawork

You can run independent multi-agent tasks in parallel to reduce completion time by firing multiple agent calls concurrently instead of sequentially. This applies to refactors, builds, and endpoint implementations where subtasks have no dependency ordering.

Model-tier routing for parallel agent orchestration assigns tasks to appropriate model tiers—LOW/Haiku for simple work, MEDIUM/Sonnet for standard work, and HIGH/Opus for complex work. This balances quality and cost across concurrently executing tasks.

You can execute long-running agent tasks in the background using optional background execution. This allows long operations to process without blocking other parallel tasks, finishing with a lightweight completion check when all tasks conclude.

Use parallel task execution for development workflows when subtasks can run without dependency ordering, such as simultaneous refactors, builds, or endpoint implementations. It maximizes throughput by avoiding sequential wait times between independent operations.

Parallel execution acts as a composable parallelism layer without persistence or verification loops, so it is intended to be wrapped by higher-level modes. It handles concurrent task firing but does not manage state verification independently.

Running parallel agent tasks requires subtasks to have no dependency ordering. It functions as a parallelism layer without persistence or verification loops, meaning it cannot manage sequential dependencies or validate task state automatically.