Chapter 13

Week 4 Day 3 - create_agent, the agent layer

Notebook.venv29 cells

Week 4 Day 3 - create_agent, the agent layer

Yesterday you built a tool loop by hand in LangGraph: a chatbot node, a tools node, a conditional edge, and an edge back again. Today you get all of that from a single function call.

create_agent is Layer 3. You hand it a model, some tools and a prompt, and it builds the agent loop for you. This is the next step along the control to convenience spine: you hand the loop to the framework and write far less code. The part that ties this whole week together is that what it builds is a LangGraph graph, the very kind you assembled by hand yesterday, and we will see that for ourselves once the agent has some tools.

In [ ]python · cell 2

python

# Imports and environment first, all in one place

from dotenv import load_dotenv
from IPython.display import Image, display
from pydantic import BaseModel, Field
from langchain.agents import create_agent
from langchain.agents.middleware import wrap_tool_call
from langchain_core.tools import tool
from langgraph.checkpoint.memory import MemorySaver
from langchain_mcp_adapters.client import MultiServerMCPClient

load_dotenv(override=True)

Part 1: the simplest agent

An agent in its plainest form is a model with a prompt. We pass the model as a string in the form provider:model, and a system prompt to set its character.

In [ ]python · cell 4

python

agent = create_agent(
    model="openai:gpt-5.4-mini",
    system_prompt="You are a helpful assistant who answers concisely.",
)

result = agent.invoke({"messages": [{"role": "user", "content": "What is the Model Context Protocol, in two sentences?"}]})
print(result["messages"][-1].content)

And its async twin: ainvoke

Everything in LangChain that can invoke can also ainvoke: same arguments, same result, awaited. You know asyncio well from earlier in the course, so there is nothing new to learn; this is simply the right way to call an agent from async code. It will matter at the end of this lab, where the browser tools only speak async.

In [ ]python · cell 6

python

result = await agent.ainvoke({"messages": [{"role": "user", "content": "In one sentence: why does async code suit agents so well?"}]})
print(result["messages"][-1].content)

Part 2: tools

To make the agent useful, we give it tools. These are the same @tool functions from Day 1. We pass them in a list, and the agent runs the whole tool loop for us.

In [ ]python · cell 8

python

@tool
def get_weather(city: str) -> str:
    """Return today's weather for a city."""
    pretend = {"London": "rainy, 14 degrees", "Rome": "sunny, 27 degrees"}
    return pretend.get(city, "clear, 20 degrees")

@tool
def get_population(city: str) -> str:
    """Return the population of a city."""
    pretend = {"London": "8.9 million", "Rome": "2.8 million"}
    return pretend.get(city, "unknown")

agent = create_agent(
    model="openai:gpt-5.4-mini",
    tools=[get_weather, get_population],
    system_prompt="You are a travel assistant. Use your tools to answer questions about cities.",
)

result = agent.invoke({"messages": [{"role": "user", "content": "What is the weather and population of Rome?"}]})
print(result["messages"][-1].content)

The reveal: it is a LangGraph graph

Now that our agent has tools, let us draw it. Because create_agent returns a compiled LangGraph graph, we can render it with exactly the same call we used yesterday. Look at the shape: a model node, a tools node, and the conditional loop between them. This is the graph you built by hand on Day 2, handed to you in one line.

In [ ]python · cell 10

python

display(Image(agent.get_graph().draw_mermaid_png()))

Part 3: memory

By default each call to invoke is a fresh start. To remember a conversation we give the agent a checkpointer, exactly as we did with the graph yesterday, and pass a thread_id so it knows which conversation we mean.

In [ ]python · cell 12

python

memory_agent = create_agent(
    model="openai:gpt-5.4-mini",
    tools=[get_weather],
    checkpointer=MemorySaver(),
)

config = {"configurable": {"thread_id": "trip-planning"}}
memory_agent.invoke({"messages": [{"role": "user", "content": "I am planning a trip to London."}]}, config=config)
result = memory_agent.invoke({"messages": [{"role": "user", "content": "What is the weather like where I am going on my trip?"}]}, config=config)
print(result["messages"][-1].content)

Part 4: structured output

When you want a typed object back rather than prose, pass a Pydantic model as response_format. The agent still does its work and uses its tools, then fills in your object. You read it from the structured_response key.

In [ ]python · cell 14

python

class CityReport(BaseModel):
    city: str = Field(description="The city name")
    weather: str = Field(description="A short weather description")
    population: str = Field(description="The population")

report_agent = create_agent(
    model="openai:gpt-5.4-mini",
    tools=[get_weather, get_population],
    response_format=CityReport,
)

result = report_agent.invoke({"messages": [{"role": "user", "content": "Give me a report on London."}]})
report = result["structured_response"]
print(report)
print("Just the weather:", report.weather)

Part 5: middleware

Middleware is a powerful feature to shape an agent's behavior. It lets you run your own code at fixed points in the loop: before the model is called, after it answers, or around each tool call.

Here is a small piece of custom middleware that prints every tool call as it happens, so you can watch the agent at work. The @wrap_tool_call decorator wraps each tool call: we log it, then call handler to let it proceed.

In [ ]python · cell 16

python

@wrap_tool_call
def log_tool_calls(request, handler):
    call = request.tool_call
    print(f"  [middleware] calling {call['name']} with {call['args']}")
    return handler(request)

watched_agent = create_agent(
    model="openai:gpt-5.4-mini",
    tools=[get_weather, get_population],
    system_prompt="You are a travel assistant. Use your tools.",
    middleware=[log_tool_calls],
)

result = watched_agent.invoke({"messages": [{"role": "user", "content": "Weather and population of London and Rome?"}]})
print("\nFinal answer:", result["messages"][-1].content)

LangChain also ships a range of ready made middleware, including SummarizationMiddleware to keep long conversations within the context window, PIIMiddleware to redact sensitive data, retry and call-limit middleware, and HumanInTheLoopMiddleware to pause for human approval.

Before Part 6: Node and Playwright

The last part of this lab uses tools that live in a separate program, an MCP server, and that program runs on Node. Two quick checks before we start.

First, Node itself; you want v22 or later. If the cell below fails, install Node with one command:

Windows, in PowerShell: winget install OpenJS.NodeJS.LTS
Mac, in Terminal: brew install node
Linux, or if neither works for you: see the guide at setup/SETUP-node.md

After installing, quit Cursor completely and start it again, then reopen this notebook and re-run the lab from the top. That full restart matters: a freshly installed Node is invisible to a notebook that was already running, and restarting just the kernel is not enough.

In [ ]python · cell 19

python

!node --version
!npx --version

Second, Playwright, Microsoft's browser automation framework. There is nothing to install: npx fetches it on demand, and it drives the copy of Chrome already on your machine. Chrome does not need to be running; Playwright launches its own.

The cell below proves the whole chain with no AI involved at all: Node runs Playwright, Playwright opens Chrome, loads Hacker News, and saves a screenshot. The first run takes a little longer while npx downloads the package.

If it complains that Chrome is not found, either install Chrome normally or run npx playwright install chrome in a terminal, then try again.

In [ ]python · cell 21

python

!npx -y playwright@latest screenshot --channel=chrome https://news.ycombinator.com playwright_check.png
display(Image("playwright_check.png"))

Part 6: an MCP server, and a real browser

Tools do not have to be Python functions you write. The Model Context Protocol is a standard way for agents to use tools that live in a separate server. LangChain loads those tools with langchain-mcp-adapters, and from the agent's point of view they are just tools like any other.

We will connect to Microsoft's Playwright MCP server, which drives the same real browser you just smoke-tested. It runs headed by default, so you can watch it work; on a Linux box with no display it quietly falls back to headless. Because these tools talk to a separate process, they are async, so we load them with await and run the agent with ainvoke, which you met in Part 1.

One adjustment for Windows

When Python launches another program, it hands that program a place to write its error messages. Inside a Jupyter notebook on Windows, that place is not a real file, and the library that launches MCP servers trips over it (a known issue in the MCP Python SDK, which only shows up in notebooks, never in ordinary Python scripts). The cell below points the MCP server's error log somewhere safe before we connect. On Mac and Linux it does nothing at all, so run it either way and carry on.

In [ ]python · cell 24

python

import sys

if sys.platform == "win32":
    import subprocess
    from functools import partial
    import langchain_mcp_adapters.sessions as mcp_sessions

    mcp_sessions.stdio_client = partial(mcp_sessions.stdio_client, errlog=subprocess.DEVNULL)
    print("Applied the Windows adjustment")
else:
    print("Not Windows, so nothing to do here")

In [ ]python · cell 25

python

client = MultiServerMCPClient({
    "playwright": {
        "transport": "stdio",
        "command": "npx",
        "args": ["-y", "@playwright/mcp@latest", "--isolated"],
    }
})

browser_tools = await client.get_tools()
print(f"Loaded {len(browser_tools)} browser tools:")
for t in browser_tools:
    print(" -", t.name)

Now we hand those browser tools to an ordinary create_agent and ask it to visit Hacker News and report what it finds. Watch the browser window open and navigate on its own.

In [ ]python · cell 27

python

browser_agent = create_agent(
    model="openai:gpt-5.5",
    tools=browser_tools,
    system_prompt="You are a web research assistant. Use the browser tools to complete the task, then report clearly.",
)

result = await browser_agent.ainvoke({"messages": [{"role": "user",
    "content": "Go to https://news.ycombinator.com and tell me the titles of the top three stories on the front page."}]})
print(result["messages"][-1].content)

Recap, and where we are heading

In one lab you have built an agent in a single line, seen that it is really a LangGraph graph, given it tools, memory and structured output, shaped it with middleware, and let it drive a real browser through an MCP server. This is the layer you will reach for most often in your own work.

Tomorrow we go one layer higher. Deep Agents take create_agent and wrap it in a full harness for work that takes many steps.

Chapter 13

Week 4 Day 3 - create_agent, the agent layer

Notebook.venv29 cells

Week 4 Day 3 - create_agent, the agent layer

Yesterday you built a tool loop by hand in LangGraph: a chatbot node, a tools node, a conditional edge, and an edge back again. Today you get all of that from a single function call.

In [ ]python · cell 2

python

# Imports and environment first, all in one place

from dotenv import load_dotenv
from IPython.display import Image, display
from pydantic import BaseModel, Field
from langchain.agents import create_agent
from langchain.agents.middleware import wrap_tool_call
from langchain_core.tools import tool
from langgraph.checkpoint.memory import MemorySaver
from langchain_mcp_adapters.client import MultiServerMCPClient

load_dotenv(override=True)

Part 1: the simplest agent

An agent in its plainest form is a model with a prompt. We pass the model as a string in the form provider:model, and a system prompt to set its character.

In [ ]python · cell 4

python

agent = create_agent(
    model="openai:gpt-5.4-mini",
    system_prompt="You are a helpful assistant who answers concisely.",
)

result = agent.invoke({"messages": [{"role": "user", "content": "What is the Model Context Protocol, in two sentences?"}]})
print(result["messages"][-1].content)

And its async twin: ainvoke

In [ ]python · cell 6

python

result = await agent.ainvoke({"messages": [{"role": "user", "content": "In one sentence: why does async code suit agents so well?"}]})
print(result["messages"][-1].content)

Part 2: tools

To make the agent useful, we give it tools. These are the same @tool functions from Day 1. We pass them in a list, and the agent runs the whole tool loop for us.

In [ ]python · cell 8

python

@tool
def get_weather(city: str) -> str:
    """Return today's weather for a city."""
    pretend = {"London": "rainy, 14 degrees", "Rome": "sunny, 27 degrees"}
    return pretend.get(city, "clear, 20 degrees")

@tool
def get_population(city: str) -> str:
    """Return the population of a city."""
    pretend = {"London": "8.9 million", "Rome": "2.8 million"}
    return pretend.get(city, "unknown")

agent = create_agent(
    model="openai:gpt-5.4-mini",
    tools=[get_weather, get_population],
    system_prompt="You are a travel assistant. Use your tools to answer questions about cities.",
)

result = agent.invoke({"messages": [{"role": "user", "content": "What is the weather and population of Rome?"}]})
print(result["messages"][-1].content)

The reveal: it is a LangGraph graph

In [ ]python · cell 10

python

display(Image(agent.get_graph().draw_mermaid_png()))

Part 3: memory

In [ ]python · cell 12

python

memory_agent = create_agent(
    model="openai:gpt-5.4-mini",
    tools=[get_weather],
    checkpointer=MemorySaver(),
)

config = {"configurable": {"thread_id": "trip-planning"}}
memory_agent.invoke({"messages": [{"role": "user", "content": "I am planning a trip to London."}]}, config=config)
result = memory_agent.invoke({"messages": [{"role": "user", "content": "What is the weather like where I am going on my trip?"}]}, config=config)
print(result["messages"][-1].content)

Part 4: structured output

In [ ]python · cell 14

python

class CityReport(BaseModel):
    city: str = Field(description="The city name")
    weather: str = Field(description="A short weather description")
    population: str = Field(description="The population")

report_agent = create_agent(
    model="openai:gpt-5.4-mini",
    tools=[get_weather, get_population],
    response_format=CityReport,
)

result = report_agent.invoke({"messages": [{"role": "user", "content": "Give me a report on London."}]})
report = result["structured_response"]
print(report)
print("Just the weather:", report.weather)

Part 5: middleware

Middleware is a powerful feature to shape an agent's behavior. It lets you run your own code at fixed points in the loop: before the model is called, after it answers, or around each tool call.

In [ ]python · cell 16

python

@wrap_tool_call
def log_tool_calls(request, handler):
    call = request.tool_call
    print(f"  [middleware] calling {call['name']} with {call['args']}")
    return handler(request)

watched_agent = create_agent(
    model="openai:gpt-5.4-mini",
    tools=[get_weather, get_population],
    system_prompt="You are a travel assistant. Use your tools.",
    middleware=[log_tool_calls],
)

result = watched_agent.invoke({"messages": [{"role": "user", "content": "Weather and population of London and Rome?"}]})
print("\nFinal answer:", result["messages"][-1].content)

Before Part 6: Node and Playwright

The last part of this lab uses tools that live in a separate program, an MCP server, and that program runs on Node. Two quick checks before we start.

First, Node itself; you want v22 or later. If the cell below fails, install Node with one command:

Windows, in PowerShell: winget install OpenJS.NodeJS.LTS
Mac, in Terminal: brew install node
Linux, or if neither works for you: see the guide at setup/SETUP-node.md

In [ ]python · cell 19

python

!node --version
!npx --version

If it complains that Chrome is not found, either install Chrome normally or run npx playwright install chrome in a terminal, then try again.

In [ ]python · cell 21

python

!npx -y playwright@latest screenshot --channel=chrome https://news.ycombinator.com playwright_check.png
display(Image("playwright_check.png"))

Part 6: an MCP server, and a real browser

One adjustment for Windows

In [ ]python · cell 24

python

import sys

if sys.platform == "win32":
    import subprocess
    from functools import partial
    import langchain_mcp_adapters.sessions as mcp_sessions

    mcp_sessions.stdio_client = partial(mcp_sessions.stdio_client, errlog=subprocess.DEVNULL)
    print("Applied the Windows adjustment")
else:
    print("Not Windows, so nothing to do here")

In [ ]python · cell 25

python

client = MultiServerMCPClient({
    "playwright": {
        "transport": "stdio",
        "command": "npx",
        "args": ["-y", "@playwright/mcp@latest", "--isolated"],
    }
})

browser_tools = await client.get_tools()
print(f"Loaded {len(browser_tools)} browser tools:")
for t in browser_tools:
    print(" -", t.name)

Now we hand those browser tools to an ordinary create_agent and ask it to visit Hacker News and report what it finds. Watch the browser window open and navigate on its own.

In [ ]python · cell 27

python

browser_agent = create_agent(
    model="openai:gpt-5.5",
    tools=browser_tools,
    system_prompt="You are a web research assistant. Use the browser tools to complete the task, then report clearly.",
)

result = await browser_agent.ainvoke({"messages": [{"role": "user",
    "content": "Go to https://news.ycombinator.com and tell me the titles of the top three stories on the front page."}]})
print(result["messages"][-1].content)

Recap, and where we are heading

Tomorrow we go one layer higher. Deep Agents take create_agent and wrap it in a full harness for work that takes many steps.