Through this guide you've progressively built:
| Example | Pattern | What It Covers |
|---|---|---|
| 1 | Hello World | Pure graph — no LLM, just nodes, edges, and state |
| 2 | Chatbot | LLM integration + conversation memory |
| 3 | ReAct | Tool use — the agent reasons, acts, and observes |
| 4 | Supervisor | Multi-agent orchestration with a supervisor routing to specialists |
| 5 | Pipeline | Sequential multi-agent pipeline (planner → searcher → analyzer → writer) |
| 8 | ReAct + RAG | Product QnA with retrieval-augmented generation + pricing tools |
| 9 | Custom Graph | Manual ReAct loop with read/write tools and StateGraph |
| 10 | Reflection | Generate → review → revise loop — self-improving output without tools |
| 11 | Multi-Agent Router | Routing pattern — classifier forwards queries to specialized agents |
These examples cover the core agentic AI patterns: tool use, planning, reflection, and multi-agent collaboration. But this is just a foundation.
LangGraph is one of several frameworks for building AI agents. Broaden your perspective by exploring alternatives:
| Framework | Focus | Link |
|---|---|---|
| LlamaIndex | Data-centric agents — great for RAG, document Q&A, and knowledge retrieval | https://www.llamaindex.ai |
| AutoGen (Microsoft) | Multi-agent conversation — agents that debate and collaborate through messages | https://github.com/microsoft/autogen |
| CrewAI | Role-based multi-agent teams with defined responsibilities | https://www.crewai.com |
| Semantic Kernel (Microsoft) | Enterprise-focused agent framework with planner and plugin architecture | https://github.com/microsoft/semantic-kernel |
Each framework has different strengths. LlamaIndex excels at data retrieval, AutoGen at multi-agent conversations, and CrewAI at defining agent roles and workflows.
Move beyond mock data and connect agents to live systems:
- RAG with real documents — Index your own PDFs, research papers, or documentation using vector stores (Chroma, Pinecone, Weaviate)
- Database integration — Connect agents to PostgreSQL, MongoDB, or other databases for live queries
- API integrations — Weather APIs, financial data, academic databases (Semantic Scholar, ArXiv), CRM systems
- Web browsing — Give agents the ability to search the web and extract information from pages
- Code execution — Let agents write and run code in sandboxed environments
Apply what you've learned to a real problem:
- Identify a repetitive workflow that involves multiple steps, decisions, or data sources
- Map it to an agentic pattern — Does it need tool use? Reflection? Multi-agent collaboration?
- Start small — Build a single-agent prototype, then expand to multi-agent if needed
- Add guardrails — System prompts, input validation, human-in-the-loop for critical decisions
- Evaluate and iterate — Use LangSmith or similar tools to trace and debug agent behavior
| Topic | Why It Matters | Resources |
|---|---|---|
| Prompt engineering | The quality of your system prompts directly determines agent behavior | OpenAI Prompt Engineering Guide |
| Evaluation | How to measure if your agent is actually working well | LangSmith, RAGAS |
| Human-in-the-loop | Pausing agents for human approval before critical actions | LangGraph HITL docs |
| Streaming | Real-time output for better UX — see each step as it happens | LangGraph Streaming |
| Deployment | Taking agents from notebooks to production APIs | LangServe, LangGraph Platform |
| Memory | Long-term memory across sessions (beyond in-memory checkpoints) | Persistent checkpointers (SQLite, PostgreSQL) |
| Safety | Preventing prompt injection, hallucination, and unintended actions | System prompt guardrails, output validation |
- LangGraph Docs — Official documentation
- LangChain Academy — Free courses on LangChain and LangGraph
- LangSmith — Observability and evaluation platform
- LlamaIndex — Data-centric agent framework
- AutoGen — Multi-agent conversation framework
- CrewAI — Role-based agent teams
- ArXiv — Latest research papers on AI agents