multi-agent-patterns

Multi-Agent Patterns

Patterns for parallel execution and multi-agent coordination in Claude Code.

---

Overview

Multi-agent patterns enable complex tasks through agent coordination, parallelization, and specialization. This document provides practical patterns for multi-agent workflows.

---

Core Patterns

1. Divide and Conquer

Problem: Task too large for single agent Solution: Split into subtasks, assign to multiple agents, aggregate results

Example: Repository-wide analysis

Coordinator: Multi-Agent Architect
├── Agent 1: Analyze /frontend
├── Agent 2: Analyze /backend
├── Agent 3: Analyze /database
└── Agent 4: Analyze /tests

Aggregation: Combine findings into report

2. Specialist Team

Problem: Task requires diverse expertise Solution: Assemble team of specialized agents

Example: Feature implementation

Team:
├── UX Designer Agent → User experience
├── Visual Designer Agent → Visual design
├── Frontend Builder Agent → UI implementation
├── API Designer Agent → Backend API
├── Security Engineer Agent → Security review
└── Testing Strategist Agent → Test coverage

Coordinator: Multi-Agent Architect

3. Pipeline Processing

Problem: Sequential stages with large data Solution: Pipeline with buffer between stages

Example: Documentation generation

Stage 1: Explore Agent → Find all files
  ↓ [Buffer: File list]
Stage 2: Analysis Agents → Extract info (parallel)
  ↓ [Buffer: Structured data]
Stage 3: Writing Agent → Generate docs
  ↓ [Buffer: Draft docs]
Stage 4: Review Agent → Polish and format

4. Fan-Out/Fan-In

Problem: Apply same operation to many items Solution: Fan out to parallel agents, fan in to aggregator

Example: Validate all modules

Fan-Out:
Source → [Splitter]
         ├→ Validator 1 (module A)
         ├→ Validator 2 (module B)
         ├→ Validator 3 (module C)
         └→ Validator 4 (module D)

Fan-In:
[Aggregator] ← All validation results
     ↓
Final Report