create_agent Basics

What changes here

In Lesson 14, we wrote the tool-call loop ourselves: ask the model, check ai_message.tool_calls, run the tool, send a ToolMessage back, ask again. That was worth doing by hand once, it's how you actually learn what's happening instead of trusting a framework blindly. But doing it by hand every time you build something would be tedious, especially once a task needs several tool calls in a row, or several different tools.

create_agent, from LangGraph (the lower-level engine LangChain's agents are built on), automates that loop. This is the first lesson where we're building something that actually deserves to be called an agent: something that can decide, on its own, to take multiple actions in order to answer you, not just respond in one shot.

Building the agent

agent = create_agent(
    model=model,
    tools=[calculator],
    system_prompt="You are a helpful assistant with access to a calculator tool.",
)

Three pieces, all familiar:

• model: the same ChatGoogleGenerativeAI object from every earlier lesson.
• tools: a list of @tool-decorated functions, same calculator from Lesson 13. You could list more than one; Lesson 15 covers that.
• system_prompt: standing instructions, same idea as the "system" message in Lesson 3's template.

create_agent returns something you still call .invoke() on, but this time it's not calling the model directly, it's running an entire loop internally: check if a tool is needed, run it if so, check again, repeat, until the model is ready to give a final answer with no more tool calls pending. The three manual rounds from Lesson 14 now happen inside one .invoke() call, and it would automatically run more rounds than three if a harder question needed them.

No memory here, on purpose

result = agent.invoke({"messages": [HumanMessage("What is 293 times 481?")]})
...
result2 = agent.invoke({"messages": [HumanMessage("What numbers did I just ask you to multiply?")]})

Notice the second call has no idea the first one happened, the agent correctly says it hasn't been asked to multiply anything yet. This is deliberate: this lesson isolates "automating the tool loop" as its own concept, before Lesson 24 adds memory on top. Each .invoke() here is still as independent as every plain model.invoke() call has been since Lesson 1.

Reading the result

result["messages"][-1].text

create_agent's result is a dictionary with a "messages" key holding the entire conversation as the agent understands it for that one call, including any tool calls and tool results that happened along the way, same shape as the messages list you built by hand in Lesson 14. The last entry is always the agent's actual final reply.