swarm

Bounded parallel execution dispatches batches of independent tasks to subagents with a concurrency cap, controlling simultaneous tool calls to manage system load. It processes parallelizable workflows like summarization or classification while returning a compact per-task summary.

Error isolation captures per-task failures without aborting the entire batch, allowing other tasks to complete. The system returns results in the original input order, providing a summary with completed outputs and specific error details for the failed items.

Dispatch a list of task descriptions to subagents with a specified concurrency limit to process items concurrently. The system returns an ordered results array containing each item's output or error, which you can iterate through to aggregate into a final report.

Ordered results processing preserves the original input order in the returned summary, regardless of the order in which concurrent subagents finish. This ensures the output array aligns correctly with your initial task list for accurate aggregation.

Using a concurrency semaphore or pool limits simultaneous subagent tool calls during parallel execution. You set a concurrency limit when dispatching the task batch, and the system caps active processes to control load before returning the summary.