Building an Agent From Scratch

What this reveals

The LangChain course's create_agent felt almost magical: hand it a model and some tools, and it handles an entire loop of deciding whether to call a tool, running it, and asking again, until it's ready to answer. It isn't magic. It's a small graph: a model node, a tool-execution node, and one conditional edge connecting them in a loop. This lesson builds that graph by hand, with the exact same primitives you already know from Beginner Lesson 7, StateGraph, MessagesState, ToolNode, tools_condition, just with a second tool added so it feels like a real, general-purpose agent instead of a toy.

Binding tools to the model

model_with_tools = model.bind_tools(tools)

bind_tools attaches the tools' schemas (their names, descriptions, and argument shapes) to every request sent to the model, so the model can respond with a request to call one of them instead of plain text. This single line is the part create_agent never shows you, it's doing this exact call internally, once, when you pass it a tools list.

The model node

def call_model(state: MessagesState) -> dict:
    response = model_with_tools.invoke(state["messages"])
    return {"messages": [response]}

Same shape as every model node since Beginner Lesson 6: read the accumulated messages, call the model, return the reply. The only difference from a plain chatbot node is that this model was bound to tools, so its reply might contain tool_calls instead of, or alongside, plain text.

Wiring the loop

builder.add_node("agent", call_model)
builder.add_node("tools", ToolNode(tools))
builder.add_edge(START, "agent")
builder.add_conditional_edges("agent", tools_condition)
builder.add_edge("tools", "agent")

Three pieces make this a loop rather than a straight line:

• ToolNode(tools): inspects the last message's tool_calls, runs the matching Python function(s), and wraps each result in a ToolMessage automatically. You never write "if the model asked for calculator, call calculator" yourself.
• tools_condition: a ready-made routing function. If the last message has pending tool calls, it routes to "tools". Otherwise, it routes to END.
• builder.add_edge("tools", "agent"): after a tool runs, control goes straight back to the model, so it can either call another tool or, once it has enough information, answer in plain text.

That's the entire ReAct loop, four lines of graph wiring. Everything create_agent does for you is contained in this shape.