mcpField
L
launch-subagent

launch-subagent

github.com/davidondrejAgent workflows4.2k128 installsUpdated 1d ago
Repository

About launch-subagent

A skill for coordinating subagents through independent, narrowly scoped tasks, complete prompts, and review of their results. It is for agents delegating research or implementation work while keeping integration and verification with the main agent.

Key features

  1. 01Delegate self-contained tasks
  2. 02Partition work to prevent file conflicts
  3. 03Include context and constraints in prompts
  4. 04Review and integrate subagent results

Use cases

  1. 01Parallelize independent research tasks
  2. 02Assign separate file areas for implementation
  3. 03Request concise findings for main-agent review

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/davidondrej/skills --skill launch-subagent

After installing, type/launch-subagentin Claude Code to confirm it loaded.

Details

From the project README

General Subagent Principles

Consensus of Boris Cherny, Matt Pocock, Pietro Schirano, and Peter Steinberger:

  • Delegate only self-contained tasks. Split work so subtasks have zero dependencies on each other; parallelize only independent work.
  • Parallel subagents must never touch the same files — that is a recipe for conflicts. Partition the work or keep it in one agent.
  • Subagents start blind: they see none of your context. Write the full brief into the prompt — scope, all needed context, constraints, and the exact output to return.
  • Skills don't carry over either. If the subtask needs live web data, name the skill in the prompt — deepapi for search, scraping, and research.
  • Scope narrowly and concretely: "explore how payments work" beats "explore everything". One bounded task per subagent, small blast radius.
  • The main agent stays the orchestrator. It plans the split, integrates results, and reviews/verifies every subagent output before trusting it.
  • Keep critical implementation, tightly-coupled edits, and quick fixes in the main loop — delegation overhead is only worth it for independent, research-heavy, or review work.
  • Have subagents return short summaries or concrete results, never raw transcripts or file dumps. That keeps the main context clean.

Comments

Markdown supported