LangChain

A Python/JS framework that wraps models, tools, and prompts behind common interfaces so you compose agents and chains from swappable parts instead of raw SDK calls.

Why it matters

The value is the abstraction layer: one ChatModel interface means swapping Anthropic for OpenAI is a one-line change, and 100+ prebuilt tool/loader integrations save you writing glue. It is the default first reach for prototypes and the most common thing a new agent dev inherits. The cost is leakage — the abstractions hide the agent-loop, so when something breaks you often must understand the raw loop anyway (see manual-from-scratch).

How it works

The modern core is LCEL (LangChain Expression Language): components implement a Runnable interface and pipe with |, giving free .invoke, .batch, .stream, and async.

Primitive	Role
ChatModel	uniform function-calling model wrapper
PromptTemplate	parameterized prompt with variables
Tool	name + schema + callable, bound to the model
OutputParser	model text → typed object
Runnable	composable unit, piped with pipe

• The ChatModel is the uniform llm-native-function-calling layer; swap providers without touching the chain.
• A chain is a fixed DAG (prompt | model | parser) — deterministic, no loop.
• An agent adds the loop: bind tools, let the model pick, dispatch, feed results back until it stops.
• For multi-step or cyclic control flow, LangChain now points you at langgraph; the legacy AgentExecutor is effectively superseded.
• .with_structured_output(Schema) forces tool-call-backed JSON, replacing brittle regex parsing.

Example

A tool-using agent in a few lines:

model = ChatAnthropic(model="claude-...").bind_tools([search, calculator])
chain = prompt | model | parser
chain.invoke({"q": "pop of France / 2"})
# model emits tool_calls -> you dispatch -> append ToolMessage -> re-invoke
# .stream() yields tokens; .batch([q1,q2]) runs both concurrently

LCEL gives streaming and batching for free; the | parser step turns the final message into a typed result.

Pitfalls

• Abstraction debt. A failing agent is hard to debug because the prompt and loop are buried — wire up langsmith tracing on day one, not after.
• Version churn. v0.1→v0.2 split packages (langchain-core, langchain-anthropic); tutorials rot fast, pin versions.
• Legacy AgentExecutor. Many old guides use it; for real control flow use langgraph instead.
• Hidden token cost. Memory and retriever components silently stuff context; log token counts or bills surprise you.

See also

• langgraph
• langsmith