multi-agent-composition_skill

Overview

This skill helps you choose and compose Claude Code components—skills, sub-agents, hooks, MCP servers, and slash commands—so you can design reliable multi-agent systems. It provides practical decision rules, composition patterns, context-management guidance, and anti-patterns to avoid when orchestrating agentic workflows. Use it to move from single prompts to scalable orchestrators with observability and control.

How this skill works

The skill inspects your use case against a decision framework that maps responsibilities to components: one-off primitives to slash commands, repeatable managed workflows to skills, parallel work to sub-agents, and external integrations to MCP servers. It evaluates context needs, prompt complexity, execution model (serial vs parallel), and observability requirements, then recommends composition patterns and implementation steps. It also highlights common anti-patterns and suggests hooks to add monitoring and lifecycle controls.

When to use it

• Deciding whether a capability should be a skill, sub-agent, slash command, or MCP integration
• Designing orchestration flows that require parallel tasks or context isolation
• Managing context size and protecting model context windows across components
• Adding observability, hooks, and lifecycle controls to agentic systems
• Scaling from single agents to multi-agent orchestration with clear separation of concerns

Best practices

• Start by mastering slash commands and prompts before building higher-level skills
• Use the Core 4 (Context, Model, Prompt, Tools) as the foundation for every component
• Assign external interactions to MCP servers and internal workflows to skills
• Use sub-agents for parallelization and context isolation; avoid nesting sub-agents
• Instrument hooks for observability, blocking, and lifecycle events from the start

Example use cases

• Convert a repeatable data-cleaning workflow into a skill that orchestrates slash commands and a sub-agent for parallel processing
• Design an orchestrator that dispatches sub-agents for batch analysis and collects results via hooks for monitoring
• Integrate an external database through an MCP server while using skills to manage business logic and retries
• Prototype one-off tasks as slash commands, then elevate to a skill only when management and repeatability are needed