delegation

Parallel task delegation assigns each independent research task, like evaluating a software library, to a separate sub-agent scoop. This architecture completes the total workload concurrently and synthesizes the results in a single turn to reduce overall execution time.

You should delegate to sub-agents when handling multi-source data collection, parallel code refactoring, or independent research. Delegation eliminates wasted context overhead from the main orchestrator's unnecessary pre-work and significantly reduces total task completion time.

Sub-agent lifecycle management provides explicit instructions for when to retain or drop sub-agents. It includes specific rules for preserving sub-agents that own active task ownership, known as sprinkles, ensuring continuous workflow tracking without premature termination.

Sandbox configuration applies best practices for setting sub-agent read and write permissions alongside command allow-lists. This configuration balances sub-agent autonomy for executing delegated tasks with strict safety constraints to protect the main orchestration environment.

Spawning one-shot ephemeral sub-agents handles quick, deterministic tasks in complete isolation. This lightweight approach executes isolated operations without retaining sub-agent state or generating follow-up context overhead for the main orchestrator.

Parallel orchestration primitives batch parallel sub-agent work together, which eliminates redundant context turns. Instead of processing individual sub-agent completion notifications sequentially, the orchestrator receives and synthesizes all compiled results simultaneously.