🔀 AgentFlow

简体中文 | English

Multi-Agent Collaboration Design Patterns with LangGraph

AgentFlow is a curated collection of battle-tested multi-agent design patterns built on LangGraph. Each pattern includes complete code, architecture diagrams, use-case analysis, and performance comparisons.

Not a framework. Not a tutorial collection. This is a design pattern reference book for multi-agent systems.

Live Demo: https://iuyup.github.io/AgentFlow

Architecture Overview

Pattern: Reflection (Iterative Self-Improvement)

Pattern: MapReduce (Parallel Fan-Out)

Screenshots & demos: See the documentation site for interactive examples and run screenshots.

Why AgentFlow?

Building multi-agent systems is hard. Not because of the tools, but because of the architecture decisions:

• When should agents loop vs. terminate?
• How do you coordinate N agents without chaos?
• When is fan-out better than sequential processing?

AgentFlow gives you proven patterns you can study, adapt, and combine — each one a complete, runnable example.

Patterns

Pattern	Description	Key Technique	Status
Reflection	Iterative self-improvement through write → review loops	Conditional looping	✅
Debate	Multi-perspective deliberation with moderator synthesis	N-party coordination	✅
MapReduce	Parallel fan-out processing with result aggregation	LangGraph Send API	✅
Hierarchical	Manager decomposes tasks → Workers execute → Manager aggregates	Nested subgraphs + Send	✅
Voting	Multiple agents independently vote, then aggregate	Broadcast fan-out	✅
GuardRail	Primary agent + safety guard checkpoint	Approve/block/redirect routing	✅
RAG-Agent	Agent decides when to retrieve from knowledge base	Conditional retrieval loop	✅
Chain-of-Experts	Task passes through specialized agents in sequence	Sequential routing	✅
Human-in-the-Loop	Key nodes wait for human confirmation	Interrupt + resume	✅
Swarm	Decentralized multi-agent collaboration	Dynamic orchestration	✅
Self-Improving	Persistent skill library + cross-task learning	Layered memory + progressive disclosure	✅

Quick Start

1. Clone & install

git clone https://github.com/iuyup/AgentFlow.git
cd AgentFlow
uv sync

2. Set up your API key

cp .env.example .env
# Edit .env and add your OpenAI API key

3. Run any pattern

python -m patterns.reflection.example
python -m patterns.debate.example
python -m patterns.map_reduce.example

4. Browse documentation

cd web
pip install -r requirements.txt
python sync_docs.py
mkdocs serve
# Visit http://localhost:8000

Project Structure

patterns/                  # Core: one directory per pattern
│   ├── reflection/        # Write → Review loop
│   ├── debate/            # N-party debate + moderator
│   ├── map_reduce/        # Parallel fan-out + reduce
│   ├── hierarchical/      # Manager → Workers → aggregation
│   ├── voting/            # Multi-agent voting + aggregation
│   ├── guardrail/         # Primary + safety checkpoint
│   ├── rag_agent/         # Agent with conditional retrieval
│   ├── chain_of_experts/  # Sequential expert routing
│   ├── human_in_the_loop/ # Human interruption
│   ├── swarm/             # Decentralized orchestration
|   └── self_improving/    # Self iteration 
├── agentflow/             # Core utilities
├── web/                   # Documentation site (MkDocs)
│   ├── docs/             # Documentation source
│   ├── mkdocs.yml        # Site configuration
│   └── sync_docs.py      # Pattern doc sync script
├── benchmarks/            # Performance comparison framework
└── docs/                  # Architecture diagrams & assets

Documentation Site

The documentation site is built with MkDocs + Material and deployed at web/:

# Local preview
cd web
pip install -r requirements.txt
python sync_docs.py    # Sync pattern docs
mkdocs serve          # Visit http://localhost:8000

# Build static site
mkdocs build

# Deploy to GitHub Pages
mkdocs gh-deploy

Requirements

• Python 3.11+
• An OpenAI API key (default model: gpt-4o-mini)

Running Tests

# Unit tests (no API key needed)
pytest patterns/

# Integration tests (requires OPENAI_API_KEY)
OPENAI_API_KEY=your-key pytest patterns/ -m "not skipif"

Design Philosophy

1. Patterns, not frameworks — Each pattern is self-contained. Copy what you need.
2. Runnable in 3 minutes — Clone, set API key, run. That's it.
3. Dual-language docs — English README + Chinese README for every pattern.
4. Real LangGraph — No abstractions over LangGraph. Learn the real API.

Contributing

See CONTRIBUTING.md for contribution guidelines.

License

MIT