v1.1 — 11 Multi-Agent Design Patterns

add self_improving agent

v1.0 — 10 Multi-Agent Design Patterns

🔀 AgentFlow v0.1.0 — First Release

10 battle-tested multi-agent design patterns built on LangGraph, each with complete code, architecture diagrams, unit tests, and dual-language documentation.

Patterns

Pattern	What it solves
🔄 Reflection	Iterative self-improvement via write → review loops
⚔️ Debate	Multi-perspective deliberation with async concurrent LLM calls
🗂️ MapReduce	Parallel fan-out with LangGraph Send API
📊 Hierarchical	Manager decomposes → Workers execute → Manager aggregates
🗳️ Voting	Independent agents vote, then aggregate results
🛡️ GuardRail	Primary agent + safety guard with approve/block/redirect
📚 RAG-Agent	Conditional retrieval with injectable retriever
🔗 Chain-of-Experts	Sequential expert routing pipeline
👤 Human-in-the-Loop	Interrupt + resume for human approval
🐝 Swarm	Decentralized collective with async message passing

Highlights

• Real async concurrency — Debate & Swarm use asyncio.gather + ainvoke for true parallel LLM calls
• Accurate LLM call counting — Built-in LLMCallCounterHandler via LangChain callbacks
• Injectable retriever — RAG-Agent accepts custom retriever functions, not just mock data
• 140 unit tests — All patterns tested with mocked LLMs, no API key needed
• Benchmark framework — Compare patterns on standardized tasks with real metrics
• Dual-language docs — English + 中文 for every pattern, deployed at iuyup.github.io/AgentFlow

Quick Start

git clone https://github.com/iuyup/AgentFlow.git
cd AgentFlow && uv sync
cp .env.example .env  # add your API key
python -m patterns.reflection.example

What's Next

• More combination examples (MapReduce + Debate + Reflection)
• Streaming support for all patterns
• LangGraph Cloud deployment guide
• Benchmark results with real LLM data