LangChain

LangChain is the most popular LLM framework in the Python and JS ecosystems. It's also the most polarizing. Both are deserved.

What it gives you

• Adapters for every model provider, vector store, and tool. One interface, many backends.
• LCEL (LangChain Expression Language). A | pipe operator for composing prompt → model → parser chains.
• A huge community. Tutorials, integrations, and Stack Overflow answers exist for almost any combination.
• Adjacent products. LangGraph for orchestration, LangSmith for tracing.

LCEL in one snippet

from langchain_core.prompts import ChatPromptTemplate
from langchain_anthropic import ChatAnthropic
from langchain_core.output_parsers import StrOutputParser

chain = (
    ChatPromptTemplate.from_template("Translate to French: {text}")
    | ChatAnthropic(model="claude-haiku-4-5")
    | StrOutputParser()
)

chain.invoke({"text": "Where is the library?"})

Pushback

• Indirection tax. Reading LangChain code often takes longer than reading the underlying SDK call would. The abstraction can hide what's actually being sent to the model.
• Churn. v0.1 → v0.2 → v0.3 broke a lot of tutorials. Pin versions and expect maintenance.
• Anti-patterns baked in. Some defaults (chat history handling, retriever wrappers) hide problems you'd rather see.

When to use it

• Prototyping a chain that touches many integrations (Slack, Google Drive, Postgres) and you want them in one place.
• A team that's already using LangSmith for tracing. LangChain integrates more cleanly there.
• Standard RAG pipelines where the abstractions match what you'd build anyway.

When to skip it

• Single-provider apps (one model, one vector store). The native SDK is fine.
• Anything where you'd benefit from full control over the prompt and tool call format.
• Production systems that need stable APIs over a 2-3 year horizon.

For agent loops specifically, LangGraph is better-shaped than the original LangChain agent abstractions.