ultrawork

Parallel agent tasks run concurrently by launching multiple independent work items at once and only serializing dependent steps. This reduces total turnaround time by executing independent subtasks simultaneously rather than sequentially.

Dependency-aware task graphs organize work items into parallel waves. Independent tasks execute concurrently, while dependent steps are serialized and only begin after their specific prerequisites complete, ensuring strict parallel-wave execution.

Yes, tasks are routed to appropriate model tiers (LOW, MEDIUM, HIGH) based on complexity. This tiered routing guidance delegates work to the most cost-effective model while maintaining quality for complex operations.

Lightweight end-state verification is required after parallel task execution, including build and typecheck validation, running affected tests, and performing manual QA to confirm implemented behavior and produce evidence-backed summaries.

Use parallel execution for implementations, investigations, and research workflows where several work items can run concurrently with shared context. It is ideal when independent subtasks exist within a larger multi-step request.

Dependent steps are sequenced by organizing tasks into parallel waves. You delegate multiple independent agent prompts simultaneously, then only sequence the dependent ones after their prerequisites complete in the previous wave.