Parallel Subagents.md

Parallel Subagents

Parallel subagents let one coordinating agent launch multiple independent sub-tasks at the same time instead of waiting for each one sequentially.

What It Does

Use parallel subagents when you have work that can be split cleanly, such as:

• comparing multiple datasets
• reviewing several code areas at once
• running validation steps in parallel
• gathering independent research results before synthesis

This reduces total wall-clock time from roughly the sum of all sub-tasks to the duration of the slowest sub-task plus coordination overhead.

Basic Pattern

spawn_subagent(
    action="spawn_parallel",
    tasks=[
        {"name": "analyzer1", "directive": "Analyze dataset A"},
        {"name": "analyzer2", "directive": "Analyze dataset B"},
        {"name": "reporter", "directive": "Summarize the findings"},
    ],
    max_concurrent=3,
    aggregate=True,
)

Key Inputs

• tasks: the list of subagent jobs to run
• max_concurrent: the concurrency cap
• aggregate: whether to return a combined summary
• fail_fast: whether the coordinator should stop on the first failure

Each task should provide a clear name and directive. Optional fields such as model overrides or tool restrictions can be added when the runtime supports them.

When To Use It

Use parallel subagents when:

• the subtasks are independent
• each subtask has a clear owner and objective
• you want faster turnaround more than strict sequential reasoning

Avoid it when:

• later steps depend tightly on earlier results
• subtasks need to share mutable state continuously
• cost or rate limits require serialized execution

Notes

• Keep directives narrow and explicit.
• Set max_concurrent to match your runtime and provider limits.
• Use aggregate=True when the parent agent needs a user-facing summary instead of raw sub-task outputs.