Your First Real LangGraph Project:

A developer has published a step-by-step guide for building a LangGraph-based customer support agent called ShopBot, which uses memory patterns, human-in-the-loop approval, and real tools to handle e-commerce inquiries. The project demonstrates how to plan a graph architecture before coding and implements nodes for order lookup, refund processing, complaint escalation, and conversation summarization. The tutorial targets beginners and emphasizes deterministic behavior using GPT-4o-mini.

For other parts of the series : Part 0 , Part 1 , Part 2 , Part 3 What this article is:Parts 0–3 gave you the architecture, the memory patterns, and the human-in-the-loop toolkit. This is where you use all three together for the first time, in one project, built step by step. Every line of code is explained. Nothing is assumed. Still very easy to follow along without reading the other parts of the series. What we’re building:A customer support agent for an e-commerce store calledShopBot. Customers can ask about their orders, request refunds, and escalate complaints. The agent remembers the conversation, uses real tools to look up order data, and pauses for human approval before processing any refund. One of the biggest mistakes beginners make is opening a blank file and starting to type. Before you write code, you need to draw the graph on paper or at least in your head. LangGraph rewards planning. Here is what ShopBot will do: That’s four nodes, two conditional edges, one interrupt call, and one summarization trigger. That's the whole agent. Write this down before you start coding. The code is just translating this picture into Python. Now let’s build it module by module, exactly as you’ve learned. Create a new file called shopbot.py. Or open a Colab notebook. Either works. Install what you need: pip install langgraph langchain langchain-openai langgraph-checkpoint-sqlite python-dotenv Create a .env file in the same folder: OPENAI API KEY=your-key-here This is always the first section. Every import your entire file needs lives here. No importing things halfway down the file. ============================================================ MODULE 1: IMPORTS & CONFIGURATION ============================================================import osimport sqlite3from typing import Annotated, Literalfrom datetime import datetimefrom dotenv import load dotenv LangChain - the AI layerfrom langchain openai import ChatOpenAIfrom langchain core.messages import HumanMessage, AIMessage, SystemMessage, BaseMessage, RemoveMessage, from langchain core.tools import tool LangGraph - the graph layerfrom langgraph.graph import StateGraph, MessagesState, START, ENDfrom langgraph.prebuilt import ToolNodefrom langgraph.checkpoint.memory import MemorySaverfrom langgraph.types import interrupt, Commandload dotenv ── The LLM ───────────────────────────────────────────────── temperature=0 means deterministic - the agent behaves consistently, which is what you want for a support bot.llm = ChatOpenAI model="gpt-4o-mini", temperature=0 Why gpt-4o-mini? It's smart enough for support tasks, cheap enough to run frequently, and fast enough that users won't notice latency. For production you might upgrade to gpt-4o for harder reasoning — but start small. The state is your agent’s working memory. Every piece of information that travels through your graph must live here. Think carefully about what your agent needs to know at each stage. ============================================================ MODULE 2: STATE ============================================================class SupportState MessagesState : MessagesState already gives us: messages: Annotated list BaseMessage , add messages We add three more fields for our specific needs: The running summary of the conversation Part 2 pattern . Starts empty. Gets written by summarize node when conversation gets long. summary: str The name of the customer, extracted early in the conversation. Used to personalise every response. Starts empty. customer name: str Tracks the current ticket category, set by the agent. Helps the human reviewer understand context during escalation. Values: "order inquiry" | "refund request" | "complaint" | "general" ticket category: str summary — you learned this in Part 2. When the conversation runs past 6 messages, the agent compresses history into this field and deletes old raw messages. Token costs stay flat. customer name — a practical field you'll see in almost every real support agent. Once the agent knows who it's talking to, it can greet them by name and personalize responses. Without it in state, each node would have to re-parse the conversation to find the name. ticket category — this is what makes the conditional routing possible. The agent sets this field, and the routing function reads it to decide whether to trigger the human approval flow. Tools are the hands of your agent. Without tools, the agent can only talk. With tools, it can actually look things up and do things . ShopBot gets three tools: ============================================================ MODULE 3: TOOLS ============================================================ ── Fake Database ──────────────────────────────────────────── In a real project, these would be database queries or API calls. For learning purposes, we use a simple Python dictionary.ORDERS DB = { "ORD-001": { "customer": "Alex", "product": "Wireless Headphones", "status": "Delivered", "amount": 89.99, "delivery date": "2025-06-10", }, "ORD-002": { "customer": "Sam", "product": "Phone Case", "status": "In Transit", "amount": 14.99, "delivery date": "Expected 2025-06-18", }, "ORD-003": { "customer": "Jordan", "product": "Laptop Stand", "status": "Processing", "amount": 45.00, "delivery date": "Expected 2025-06-20", },}@tooldef lookup order order id: str - str: """Look up the details of a customer's order by order ID. Use this when the customer provides an order number and wants to know the status, product name, or delivery date of their order. Args: order id: The order ID string, e.g. 'ORD-001' Returns: A formatted string with full order details, or an error message if the order is not found. """ order = ORDERS DB.get order id.upper if not order: return f"No order found with ID '{order id}'. Please double-check the order number." return f"Order {order id.upper }: {order 'product' } | " f"Status: {order 'status' } | " f"Amount: ${order 'amount' :.2f} | " f"Delivery: {order 'delivery date' }" @tooldef check refund eligibility order id: str - str: """Check whether an order is eligible for a refund. Use this BEFORE processing any refund request. An order is eligible for a refund only if its status is 'Delivered'. Orders Fthat are 'In Transit' or 'Processing' cannot be refunded yet. Args: order id: The order ID string, e.g. 'ORD-001' Returns: A string stating whether the order is eligible and why. """ order = ORDERS DB.get order id.upper if not order: return f"Cannot check refund: order '{order id}' not found." if order "status" == "Delivered": return f"Order {order id.upper } IS eligible for a refund. " f"Product: {order 'product' }, Amount: ${order 'amount' :.2f}. " f"Proceed to refund processing." else: return f"Order {order id.upper } is NOT eligible for a refund yet. " f"Current status: {order 'status' }. Refunds are only available " f"for delivered orders." @tooldef process refund order id: str, reason: str - str: """Process a refund for a delivered order. IMPORTANT: This tool actually issues the refund. It should only be called AFTER human approval has been obtained. Never call this tool without prior confirmation. Args: order id: The order ID to refund reason: The customer's stated reason for the refund Returns: A confirmation string with the refund reference number. """ order = ORDERS DB.get order id.upper if not order: return f"Refund failed: order '{order id}' not found." In a real system, this would hit your payments API. refund ref = f"REF-{order id.upper }-{datetime.now .strftime '%H%M%S' }" return f"Refund APPROVED and PROCESSED. Reference: {refund ref}. " f"${order 'amount' :.2f} will be returned to the original payment method " f"within 3–5 business days. Reason logged: '{reason}'." ── Collect tools and bind to LLM ─────────────────────────── All three tools in one list.tools = lookup order, check refund eligibility, process refund llm with tools = the LLM that KNOWS about the tools and can decide to call them. This is what we use inside agent node.llm with tools = llm.bind tools tools tool node = the pre-built node that EXECUTES whatever tool the LLM chose. This is what we register in Module 6.tool node = ToolNode tools Look at the docstring for process refund. It says "This tool actually issues the refund. It should only be called AFTER human approval." The LLM reads this. It informs the LLM's decision-making. Write docstrings as if you're giving instructions to a smart intern who doesn't know your business rules yet — because that's exactly what you're doing. This is where the work happens. Each node receives state, does something, and returns a dictionary of updates. Nothing else. ShopBot has three nodes you write yourself, plus tool node from Module 3 pre-built . ============================================================ MODULE 4: NODES ============================================================ ── The System Prompt ──────────────────────────────────────── Written once, used in every call to the LLM from agent node. This is the personality and rulebook of your agent.SYSTEM PROMPT = """You are ShopBot, a friendly and professional customer support \agent for an e-commerce store.Your capabilities:- Look up order details using the lookup order tool- Check if an order qualifies for a refund using check refund eligibility- Process approved refunds using the process refund toolYour rules:- Always greet the customer by name once you know it- Always check refund eligibility BEFORE attempting to process a refund- For refund requests, set ticket category to "refund request" in your reasoning- Be empathetic, clear, and concise- If you cannot help, offer to escalate to a human agentImportant: The process refund tool requires prior human approval. Do not call it \unless the conversation shows that a human has already approved the refund.""" ── Node 1: agent node ──────────────────────────────────────def agent node state: SupportState - dict: """The brain of the operation. Reads state, calls the LLM, and decides whether to use a tool, give a final answer, or do something else. This node handles two cases: 1. Normal conversation - just call the LLM and respond 2. Long conversation - if a summary exists, prepend it so the LLM has context without seeing all the raw messages """ Part 2 pattern: check for an existing summary summary = state.get "summary", "" if summary: Build context: system prompt + compressed history + recent messages system with summary = SystemMessage content=f"{SYSTEM PROMPT}\n\nSummary of conversation so far:\n{summary}" messages to send = system with summary + state "messages" else: No summary yet - full history is short enough to send as-is system msg = SystemMessage content=SYSTEM PROMPT messages to send = system msg + state "messages" Call the LLM. It sees tools and can choose to call one. response = llm with tools.invoke messages to send Detect ticket category from the response for routing purposes. A smarter version would have the LLM explicitly set this - for now, we scan for keywords. content lower = response.content.lower if response.content else "" updates: dict = {"messages": response } if "refund" in content lower or hasattr response, "tool calls" and any "refund" in str tc .lower for tc in response.tool calls or : updates "ticket category" = "refund request" return updates ── Node 2: review refund ───────────────────────────────────def review refund state: SupportState - dict: """The human approval gate. Pauses execution, shows the pending refund details to a human agent, and waits for their decision. This implements the Part 3 interrupt pattern. Execution stops here until someone calls graph.invoke Command resume=... , config . Three outcomes the human can choose: - "approve" → let the refund tool call proceed unchanged - "reject" → cancel the refund, send a message to the customer - "escalate" → hand the entire ticket to a human support agent """ last message = state "messages" -1 Find the refund-related tool call in the last AI message. We look for process refund specifically - the "real action" tool. refund tool call = None if hasattr last message, "tool calls" : for tc in last message.tool calls: if "refund" in tc "name" .lower : refund tool call = tc break Surface the context to the human reviewer via interrupt . Everything in this dict is what the human sees before deciding. human decision = interrupt { "message": "⚠️ Refund approval required", "customer name": state.get "customer name", "Unknown" , "tool being called": refund tool call "name" if refund tool call else "refund tool", "arguments": refund tool call "args" if refund tool call else {}, "conversation summary": state.get "summary", "No summary yet" , "options": "approve", "reject", "escalate" , } ── Handle the human's decision ───────────────────────── if human decision == "approve": Do nothing to state - let tool node execute the tool call as-is return {} elif human decision == "reject": Cancel the tool call. The LLM will see a ToolMessage explaining why, and generate a polite response to the customer. from langchain core.messages import ToolMessage return { "messages": ToolMessage content= "Refund request was reviewed and declined by our support team. " "Please inform the customer politely and offer alternatives." , tool call id=refund tool call "id" if refund tool call else "unknown", } elif human decision == "escalate": Signal escalation - in a real system you'd open a ticket, ping Slack, or transfer to a live agent queue. from langchain core.messages import ToolMessage return { "messages": ToolMessage content= "This ticket has been escalated to a senior support agent. " "Inform the customer that a human agent will contact them " "within 2 business hours." , tool call id=refund tool call "id" if refund tool call else "unknown", } Fallback - treat as approve return {} ── Node 3: summarize node ──────────────────────────────────def summarize node state: SupportState - dict: """Triggered when the conversation exceeds 6 messages. Compresses the full message history into a short summary, then deletes old raw messages. This is the rolling summary pattern from Part 2. The summary grows richer turn by turn. Token costs stay nearly flat no matter how long the conversation runs. """ existing summary = state.get "summary", "" if existing summary: Extend the existing summary with new messages summary instruction = f"Current summary:\n{existing summary}\n\n" "Extend this summary with the new messages above. " "Keep it under 5 sentences. Focus on: the customer's name, " "their issue, any orders mentioned, and what actions were taken." else: First time summarising summary instruction = "Summarise this customer support conversation in under 5 sentences. " "Include: the customer's name if mentioned , their issue, " "any order numbers discussed, and what actions were taken so far." messages = state "messages" + HumanMessage content=summary instruction response = llm.invoke messages Plain llm, no tools needed here Delete all but the 2 most recent messages. The summary now holds everything that was in the deleted messages. messages to delete = RemoveMessage id=m.id for m in state "messages" :-2 return { "summary": response.content, "messages": messages to delete, } Routing functions are the decision-makers. They look at state and return a string. That’s all they do. They never modify state — that’s the nodes’ job. ShopBot needs two routing functions: ============================================================ MODULE 5: EDGES & ROUTING ============================================================def route after agent state: SupportState - Literal "review refund", "tools", "summarize node", " end " : """Called after agent node runs. Decides what happens next. Four possible routes: 1. The LLM wants to call process refund → must go through human review first 2. The LLM wants to call any other tool → go directly to tool node 3. The LLM gave a plain text answer AND the conversation is long → summarise 4. The LLM gave a plain text answer and conversation is short → we're done """ last message = state "messages" -1 has tool calls = hasattr last message, "tool calls" and bool last message.tool calls if has tool calls: Check if ANY of the tool calls is the sensitive process refund tool tool names = tc "name" for tc in last message.tool calls if "process refund" in tool names: return "review refund" → Pause for human approval first return "tools" → Safe tool, run it directly No tool call - the LLM gave a plain response. Check if the conversation is long enough to need summarisation. if len state "messages" 6: return "summarize node" return " end " → Conversation turn is completedef route after review state: SupportState - Literal "tools", "agent node" : """Called after review refund runs i.e., after the human has decided . Two routes: 1. Human approved or escalated → run the tool tool node handles the call 2. Human rejected → the review node already added a ToolMessage cancelling the tool call, so skip tool node and go back to agent node to respond """ last message = state "messages" -1 If the last message is a ToolMessage, the review node cancelled the call. Go back to agent node so it can generate a customer-facing response. from langchain core.messages import ToolMessage if isinstance last message, ToolMessage : return "agent node" Otherwise, the review node returned {} approved - proceed to tools. return "tools" This is construction day. All the pieces exist. Now you wire them together. Follow the five-step sequence every time: Initialize → Register → Entry → Edges → Compile. ============================================================ MODULE 6: GRAPH ASSEMBLY ============================================================ ── Step 1: Initialize ──────────────────────────────────────graph builder = StateGraph SupportState ── Step 2: Register All Nodes ────────────────────────────── Format: add node "string name", function The string name is what you use in every edge definition below.graph builder.add node "agent node", agent node graph builder.add node "tools", tool node Pre-built from Module 3graph builder.add node "review refund", review refund graph builder.add node "summarize node", summarize node ── Step 3: Set Entry Point ───────────────────────────────── The first node that runs when a user sends a message.graph builder.add edge START, "agent node" ── Step 4: Wire the Edges ────────────────────────────────── After agent node: conditional - depends on what the LLM decidedgraph builder.add conditional edges "agent node", Source route after agent, Router function from Module 5 { "review refund": "review refund", Refund tool → human review first "tools": "tools", Other tools → run directly "summarize node": "summarize node", Long conversation → summarise " end ": END, Plain answer → done } After review refund: conditional - depends on human's decisiongraph builder.add conditional edges "review refund", route after review, { "tools": "tools", Approved → execute the tool "agent node": "agent node", Rejected → back to agent to respond } After tools run: always go back to agent node the ReAct loop - agent sees tool result, decides what to do next graph builder.add edge "tools", "agent node" After summarization: conversation turn is donegraph builder.add edge "summarize node", END ── Step 5: Compile ───────────────────────────────────────── Using MemorySaver for development. For production, swap this one line to SqliteSaver or PostgresSaver.memory = MemorySaver shopbot = graph builder.compile checkpointer=memory If you’re in a Colab or Jupyter notebook, run this to see a diagram of your graph: python from IPython.display import display, Imagefrom langchain core.runnables.graph import MermaidDrawMethoddisplay Image shopbot.get graph .draw mermaid png draw method=MermaidDrawMethod.API This is one of the best features of LangGraph for learners — you can see exactly what you built, drawn as a real flowchart. If something looks wrong in the diagram, something is wrong in your edges. This is where you run the agent. For a script, it’s the if name == " main " block. For an API server, it's your route handler. The pattern is the same either way. ============================================================ MODULE 7: ENTRYPOINT ============================================================def run shopbot : """ Interactive command-line session with ShopBot. Demonstrates: multi-turn conversation, tool use, and human-in-the-loop. """ print "=" 55 print " ShopBot - Customer Support Agent" print " Powered by LangGraph" print "=" 55 print "Type your message below. Type 'exit' to quit." print "Type 'state' to inspect what ShopBot currently remembers.\n" One config per session. thread id is the session key - same ID = same memory thread. Change the ID to start a completely fresh conversation. config = {"configurable": {"thread id": "customer-session-001"}} while True: user input = input "You: " .strip if not user input: continue if user input.lower == "exit": print "ShopBot: Thank you for contacting support. Have a great day " break ── Debug: inspect current state ──────────────────── if user input.lower == "state": snapshot = shopbot.get state config print "\n DEBUG Current State:" print f" Messages in state : {len snapshot.values.get 'messages', }" print f" Customer name : {snapshot.values.get 'customer name', ' not set ' }" print f" Ticket category : {snapshot.values.get 'ticket category', ' not set ' }" print f" Summary : {snapshot.values.get 'summary', ' none yet ' }" print f" Next node s : {snapshot.next}\n" continue ── Normal message: invoke the graph ───────────────── result = shopbot.invoke {"messages": HumanMessage content=user input }, config=config, ── Check if graph paused for human approval ───────── This is how you detect that interrupt was called inside review refund. while " interrupt " in result: interrupt data = result " interrupt " 0 .value print "\n" + "=" 55 print " HUMAN APPROVAL REQUIRED" print "=" 55 print f" Customer : {interrupt data.get 'customer name', 'Unknown' }" print f" Action : {interrupt data.get 'tool being called', 'refund' }" print f" Arguments : {interrupt data.get 'arguments', {} }" print f" Context : {interrupt data.get 'conversation summary', 'N/A' }" print "=" 55 print "Options: a Approve r Reject e Escalate" human choice = input "Your decision: " .strip .lower if human choice == "a": resume value = "approve" elif human choice == "r": resume value = "reject" elif human choice == "e": resume value = "escalate" else: print "Invalid choice. Defaulting to reject." resume value = "reject" Resume the graph with the human's decision. Command resume=... answers the pending interrupt call. result = shopbot.invoke Command resume=resume value , config=config, ── Print the agent's final response ───────────────── last message = result "messages" -1 print f"\nShopBot: {last message.content}\n" if name == " main ": run shopbot Here is a full example session so you know what to expect when you run this. ======================================================= ShopBot — Customer Support Agent Powered by LangGraph=======================================================Type your message below. Type 'exit' to quit.Type 'state' to inspect what ShopBot currently remembers.You: Hi, my name is Alex and I need help with order ORD-001ShopBot: Hi Alex I'd be happy to help you with order ORD-001.I can see your order:- Product: Wireless Headphones- Status: Delivered on 2025-06-10- Amount: $89.99What can I help you with regarding this order?You: I want to return it and get a refundShopBot: I've checked your order ORD-001 and it's eligible for a refundsince it's been delivered. Let me process that for you right away.======================================================= HUMAN APPROVAL REQUIRED======================================================= Customer : Alex Action : process refund Arguments : {'order id': 'ORD-001', 'reason': 'Customer requested return'} Context : No summary yet=======================================================Options: a Approve r Reject e EscalateYour decision: aShopBot: Great news, Alex Your refund has been successfully processed.Refund Reference: REF-ORD-001-143022Amount: $89.99 will be returned to your original payment method within3–5 business days.Is there anything else I can help you with?You: state DEBUG Current State: Messages in state : 6 Customer name : not set Ticket category : refund request Summary : none yet Next node s : You: exitShopBot: Thank you for contacting support. Have a great day Let’s be explicit about what every part of this project maps to in the series: From Part 1 Architecture : The entire seven-module file structure. Every module in exactly the right order. StateGraph, add node, add conditional edges, compile — all used correctly. From Part 2 Memory : The summary field in state. The summarize node function. The rolling summary pattern — new messages get compressed, old ones deleted, token costs stay flat. The MemorySaver checkpointer keeping conversation history alive across turns. From Part 3 Human-in-the-Loop : The review refund node using interrupt to pause execution. The Command resume=... call in the entrypoint to resume. The while " interrupt " in result loop to handle cases where multiple approvals might be needed in a single turn. The three-option approval flow approve / reject / escalate . New in this project: Real tools with @tool. The ToolNode pre-built node. Binding tools to the LLM with bind tools. The ReAct loop tools → agent node → tools → ... . Conditional routing that treats different tools differently safe tools go straight to tool node; sensitive tools go through review refund first . The project above is deliberately complete but minimal. Here are three specific things to add next, in order of difficulty: Level 1 — Better category detection: Right now, ticket category is set by keyword scanning in agent node. A cleaner approach is to ask the LLM to explicitly categorize the ticket as part of its response, using structured output. Look up with structured output in LangChain docs. Level 2 — SQLite persistence: Swap MemorySaver for SqliteSaver one line change in Module 6 . Add a simple loop that asks "same session or new?" at startup. Your conversations now survive restarts. You learned exactly how to do this in Part 2. Level 3 — Add a complaint escalation path: Add a fourth node called escalate node that fires when ticket category == "complaint" and the customer emotion is "FRUSTRATED". The node calls interrupt to notify a human agent. Wire it in as a new branch off route after agent. Everything you need for this is already in Part 3. Everything new this project introduced, in one place: ToolNode tools — pre-built node that executes tool calls from the LLM's last message. Takes your tools list, handles everything else automatically. llm.bind tools tools — connects your tool list to the LLM. After binding, the LLM can choose to call any tool by returning a special tool calls field in its response instead of plain text. @tool — decorator that turns a Python function into an LLM-callable tool. The function's docstring is what the LLM reads to decide when to use it. hasattr message, "tool calls" and message.tool calls — the standard pattern for checking if the LLM's last message contains a tool call. Always check hasattr first before accessing the attribute. while " interrupt " in result — the loop pattern in Module 7 that handles one or more pending interrupt calls. Each Command resume=... answers one interrupt and potentially triggers the next. shopbot.get state config — inspect the full current state at any point. .values gives you the state dict. .next tells you what would run next. Invaluable for debugging. Reading Parts 1, 2, and 3 gave you the map. This project is the first step into the territory. The gap between understanding a concept and writing the code for it is real, and it’s filled by exactly this kind of project — something small enough to hold in your head completely, but complete enough to use every tool in your toolkit. You have a checkpointer. You have real tools. You have a human approval gate. You have a summarization trigger. Every major concept from the series, working together in one coherent agent. Build this. Break it deliberately. Fix the break. Then extend it. That’s the only path from reading articles to writing production code. Next in the series: Multi-Agent Systems — splitting ShopBot into specialist sub-agents a Refund Agent, an Order Agent, a Complaints Agent orchestrated by a Supervisor, and connecting them using LangGraph’s subgraph pattern. For other parts of the series : Part 0 , Part 1 , Part 2 , Part 3 Your First Real LangGraph Project: https://pub.towardsai.net/your-first-real-langgraph-project-ac5eb00f923a was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.