Parallel Agent Workflows

Parallelization reduces wall-clock time when subtasks are independent. Examples include analyzing several documents or asking specialist evaluators to score the same candidate answer.

Do not parallelize dependent work

If task B needs task A's result, running them together creates retries or inconsistent assumptions.

Bound concurrency

Parallel agents can multiply API usage instantly. Use concurrency limits and per-run budgets.

Aggregate explicitly

Define how partial results are merged. A final aggregator should know which outputs succeeded, failed, or timed out.