AI Agent Frameworks • Module A: LangChain & LangGraphLesson 1: LangChain — Chains, Runnables & LCEL
LangChain is the most widely used AI application framework. This lesson covers what actually matters for building production systems: LCEL pipe syntax, chain composition, and custom runnables — skipping the cruft.
LangChain is a composition library. Its value is in standardizing how LLMs, prompts, retrievers, tools, and parsers connect to each other — so you can swap components without rewriting the integration code. LCEL (LangChain Expression Language) is the composable core of modern LangChain.
LCEL is LangChain's pipe syntax: compose components with | operator. The result is a Runnable that can be invoked, streamed, or batched uniformly.
from langchain_anthropic import ChatAnthropic
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
model = ChatAnthropic(model="claude-haiku-4-5-20251001")
# LCEL chain: prompt | model | parser
chain = (
ChatPromptTemplate.from_template("Translate to French: {text}")
| model
| StrOutputParser()
)
# Invoke
result = chain.invoke({"text": "Hello, how are you?"})
print(result) # → "Bonjour, comment allez-vous?"
# Stream token by token
for chunk in chain.stream({"text": "The weather is nice today"}):
print(chunk, end="", flush=True)
# Batch (parallel)
results = chain.batch([
{"text": "Good morning"},
{"text": "Good night"},
{"text": "See you later"},
])| Use LangChain when | Use raw SDK when |
|---|---|
| Building a RAG pipeline (retrievers are first-class) | Simple 1–2 step LLM calls |
| Need streaming + batching + caching consistently | Maximum control and minimal dependencies |
| Composing 5+ steps that need easy observability | Learning how LLMs work at the protocol level |
| Integrating with LangSmith for tracing | The LangChain abstraction adds more confusion than value |
# Set these env vars — LangChain auto-instruments all calls import os os.environ["LANGCHAIN_TRACING_V2"] = "true" os.environ["LANGCHAIN_API_KEY"] = "your-langsmith-key" os.environ["LANGCHAIN_PROJECT"] = "my-rag-app" # Now every chain.invoke() is automatically traced: # - Input/output for every step # - Token counts and latency # - Error traces with full context # All visible in the LangSmith UI at smith.langchain.com
In the next lesson, we move from stateless chains to stateful agents using LangGraph — where the flow of execution depends on the model's decisions.