Exercise to practice LangChain Tools and Agents (LC<0.3)

exercise_langchain02_agents.md

Practice: LangChain Agents

Setup

Install dependencies:

#
# Tip: Use a virtual environment to keep this project's dependencies isolated from your system Python and other projects.
#
!pip install "langchain<0.3" "langchain-core<0.3" "langchain-community<0.3" "langchain-openai<0.2"

from langchain.chat_models import ChatOpenAI
from langchain.agents import Tool, initialize_agent
from langchain.memory import ConversationBufferWindowMemory

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)

---

Iteration 1: Give the agent tools

Below are two tools already defined for you.

def calculator(expression: str) -> str:
    """Evaluates a math expression, e.g. '12 * 4'."""
    return str(eval(expression))

def word_counter(text: str) -> str:
    """Counts the number of words in a text."""
    return str(len(text.split()))

tools = [
    Tool(name="Calculator", func=calculator, description="Useful for solving math expressions."),
    Tool(name="WordCounter", func=word_counter, description="Useful for counting words in a text."),
]

Your task:

1. Create an agent with initialize_agent() using agent="chat-conversational-react-description".
2. Run it with a question that requires the Calculator tool (e.g. "What is 245 * 12?").

💡 Solution

agent = initialize_agent(
    tools=tools,
    llm=llm,
    agent="chat-conversational-react-description",
    verbose=True,
)

agent.run("What is 245 * 12?")

---

Iteration 2: Add memory

Your task: Add a ConversationBufferWindowMemory (k=3) to your agent so it remembers previous turns. Ask two related questions, one after the other (e.g. "My name is Sam." then "What's my name?").

💡 Solution

memory = ConversationBufferWindowMemory(
    memory_key="chat_history", k=3, return_messages=True
)

agent = initialize_agent(
    tools=tools,
    llm=llm,
    agent="chat-conversational-react-description",
    memory=memory,
    verbose=True,
)

agent.run("My name is Sam.")
agent.run("What's my name?")

---

Bonus 1: A third tool, combined

Add a new tool that looks up a "weather" from a fixed dictionary:

def get_weather(city: str) -> str:
    """Returns the weather for a given city."""
    fake_weather = {"Madrid": "Sunny, 30°C", "London": "Rainy, 15°C"}
    return fake_weather.get(city, "No data for that city.")

tools.append(Tool(name="Weather", func=get_weather, description="Useful for checking the weather in a city."))

Ask the agent a question that requires both the Weather and Calculator tools in the same query (e.g. "What's the weather in Madrid, and what is 15 * 3?").

💡 Solution

tools = [...] # make sure to pass all the tools needed (including the new one)

agent = initialize_agent(
    tools=tools,
    llm=llm,
    agent="chat-conversational-react-description",
    memory=memory,
    verbose=True,
)

agent.run("What's the weather in Madrid, and what is 15 * 3?")

Check the verbose output — the agent should call Weather and Calculator separately, then combine both results in its final answer.

---

Bonus 2: Read the agent's reasoning

Run any query with verbose=True and look at the printed trace (the Thought / Action / Action Input / Observation steps).

Check which tool the agent picked and why — based only on the trace, not on your own guess.

💡 Solution

There's no single "correct" answer here — check that the agent behaves as expected.

---

Bonus 3: Handle tool errors gracefully

This tool fails on purpose sometimes:

import random

def flaky_tool(query: str) -> str:
    """A tool that randomly fails."""
    if random.random() < 0.5:
        raise ValueError("Tool temporarily unavailable")
    return "Success!"

tools.append(Tool(name="FlakyTool", func=flaky_tool, description="Useful when asked to 'test the flaky tool'."))

Your task: Re-create the agent with handle_parsing_errors=True and a max_iterations limit, so it doesn't crash or loop forever when the tool fails. Test by asking it to "test the flaky tool" a few times.

💡 Solution

agent = initialize_agent(
    tools=tools,
    llm=llm,
    agent="chat-conversational-react-description",
    memory=memory,
    verbose=True,
    handle_parsing_errors=True,
    max_iterations=3,
)

agent.run("Test the flaky tool.")

---

Additional Bonus: Use a retriever as a tool

For a more complex bonus, you can provide a retrievar as a tool.