What should a framework simplify, and what must stay visible?

Your application needs to send messages, call tools, and inspect responses. A framework can give those operations a common shape. But if an answer is wrong, you still need to know what reached the model and what came back.

Before you begin: Understand a direct SDK request and Python environment setup.

Start with the request you already understand

LangChain provides model interfaces and an agent harness, including the current create_agent entry point. Its agents build on LangGraph's orchestration. LangSmith is a separate service for tracing and evaluation. These names refer to different responsibilities; installing one does not automatically configure every other capability.

The current documentation was reviewed on September 10, 2026. Older tutorials using legacy chain classes or memory abstractions may target a different API generation. Match documentation to the installed packages instead of mixing imports from several years of examples.

Make one model call first

In an isolated Python environment, install langchain and the provider integration you intend to use. For the OpenAI integration, the current documentation uses the langchain[openai] extra. Record installed versions with python -m pip freeze after setup.

python -m pip install 'langchain[openai]'

This integration example requires OPENAI_API_KEY and OPENAI_MODEL in the process environment. Choose a model actually available to your account. It was checked against the current interface but was not executed against a paid provider during review.

import os
from langchain.chat_models import init_chat_model

model = init_chat_model(
    os.environ['OPENAI_MODEL'],
    model_provider='openai',
    timeout=30,
    max_retries=0,
)
reply = model.invoke([
    {'role': 'system', 'content': 'Use only the supplied notice.'},
    {'role': 'user', 'content': (
        'Notice: The class is in Room 4. '
        'Question: Which room should I go to?'
    )},
])
print(reply.content_blocks)

The input is a sequence of messages. The result is an AI message object, not necessarily a plain string. Content can include typed blocks, tool calls, and metadata depending on the model. Inspect the actual result before writing an output parser that assumes every response has one text field.

Know what the common interface does not erase

Providers differ in supported parameters, tool calling, structured output, streaming events, and error behavior. A common method name such as invoke reduces adapter work; it does not prove that switching providers leaves every behavior unchanged.

In particular, do not set a sampling parameter merely because another model accepted it. Check compatibility and evaluate the new provider's response structure. Keep provider-specific decisions visible in configuration instead of hiding them in a generic prompt file.

Compose only the steps you need

For a grounded answer, the application may validate input, select allowed evidence, format messages, call the model, and inspect the result. A single Python function can be enough at first. Framework composition becomes useful when it clarifies repeated interfaces or orchestration, not when it adds layers around a one-line operation.

If you add an agent, identify who chooses the next step. If you add persistence, identify where it is stored and how it is scoped. If you add tracing, understand what request content the tracing system receives. None of these should happen accidentally because a tutorial enabled every integration.

Inspect a failure at the boundary

Replace the question with one the notice cannot answer. If the reply invents a time, save the exact messages and response for your local evaluation, excluding credentials. Then verify that the notice reached the request. This distinguishes prompt construction errors from model behavior.

Next, we will add a tool and compare a fixed sequence with a model-controlled loop.

References

Use the current LangChain overview and model interface guide. They document the abstraction; your provider's reference remains relevant for supported options.

Continue to the next lesson.