How the LangChain stack fits together

Five pages covering the pieces you actually touch in the LangChain 1.x stack: one per package for building an agent, and two for retrieval. Every entry says what the class or function defines or does, links to the official API page (opens in a new tab), and ends with a runnable example whose printed output was verified against the pinned versions.

How the three parts relate

 1              you write                          you run
 2                 │                                  │
 3   langchain     │  create_agent(model, tools, middleware=…, response_format=…)
 4   (agent layer) │        │ builds a StateGraph and calls .compile()
 5                 ▼        ▼
 6   langgraph     │  CompiledStateGraph.invoke / stream / get_state
 7   (runtime)     │        │ runs nodes, merges state, checkpoints, pauses on interrupt()
 8                 ▼        ▼
 9   langchain-core│  BaseChatModel.invoke(list[BaseMessage]) -> AIMessage
10   (types)       │  BaseTool.invoke(tool_call) -> ToolMessage
11                 │  HumanMessage / AIMessage / ToolMessage / SystemMessage
12                 ▼
13   provider pkg  │  langchain-anthropic, langchain-openai, … subclass BaseChatModel

The dependency arrow points down and only down. Verified from the installed packages' metadata and by import tracing:

• langchain 1.4.0 requires langgraph<1.3.0,>=1.2.11 and langchain-core<2.0.0,>=1.6.0.
• langgraph 1.2.11 requires langchain-core<2,>=1.4.7. It does not require langchain, and importing it does not import langchain.
• langchain-core 1.6.2 requires neither. Importing every one of its modules leaves langgraph and langchain absent from sys.modules.
• langchain-text-splitters 1.1.2 requires only langchain-core>=1.2.31, and importing it leaves langgraph and langchain absent too. It sits beside the stack, not in it.

What that means in practice:

• An agent is a graph. create_agent (langchain) returns a CompiledStateGraph (langgraph). Everything the langgraph file says about invoke, stream, get_state, checkpointers and interrupt applies to an agent unchanged, and agent.get_graph().draw_mermaid() shows you the model ⇄ tools loop it built.
• The graph's state is a list of langchain-core messages. AgentState.messages (langchain) is MessagesState.messages (langgraph) is list[AnyMessage] merged by add_messages, and the items are HumanMessage/AIMessage/ToolMessage (langchain-core).
• Models and tools are langchain-core objects. create_agent calls BaseChatModel.bind_tools and runs BaseTool.invoke(tool_call); the provider package supplies the concrete model class.
• Middleware is the seam between the layers. AgentMiddleware hooks (langchain) receive the langgraph state and a ModelRequest holding langchain-core messages and tools.
• Retrieval is langchain-core plus one sibling package. Document, Embeddings, VectorStore and BaseRetriever are core types — a retriever is a Runnable, so it drops into a | chain or a graph node like any other — and chunking is langchain-text-splitters, which depends on core alone. Neither touches langgraph or langchain.

Which page to open

Each page has the same shape: pinned version → what the package is → an Index table (name, what it defines/does, most-used members) → one section per name with the official link, a concrete description, a verified example, and for graphs the Mermaid diagram the code itself printed.

Reproducing the examples

1python3 -m venv .venv && source .venv/bin/activate
2pip install "langchain==1.4.0" "langchain-core==1.6.2" "langgraph==1.2.11" "langchain-anthropic==1.7.1" \
3            "langgraph-checkpoint-sqlite==3.1.1" "langchain-text-splitters==1.1.2" "numpy==2.5.2"

langchain-text-splitters 1.1.2 is the version pip resolves beside core 1.6.2 with no pin. numpy is declared by none of these packages, but InMemoryVectorStore and the fake embeddings import it lazily and fail without it; 2.5.2 is what resolved.

All examples run offline (they use langchain_core's GenericFakeChatModel or a five-line BaseChatModel subclass) except one, marked The one live example in the langchain reference, which needs ANTHROPIC_API_KEY and skips itself without it. Verified on Python 3.13.7 — the three stack pages on 2026-09-05, the two retrieval pages on 2026-09-06.

Old-tutorial warnings, collected

• langgraph.prebuilt.create_react_agent → deprecated since LangGraph 1.0 (warning class LangGraphDeprecatedSinceV10, removal planned for 2.0). Use langchain.agents.create_agent.
• AgentExecutor, initialize_agent, LLMChain, ConversationChain and the other Chain classes → moved to the separate langchain-classic package (1.0.8); importing them from langchain.agents raises ImportError. A tutorial using them predates 1.0.
• langchain-mcp-adapters → its tool conversion was ported into langchain.mcp (MCPAdapter) in langchain 1.4.0; needs the langchain[mcp] extra.
• langchain.text_splitter, langchain.vectorstores, langchain.document_loaders, langchain.schema → ModuleNotFoundError in langchain 1.4.0. The splitters are langchain_text_splitters; Document, Embeddings, VectorStore and BaseRetriever are langchain_core.*; concrete loaders and stores are their own packages (langchain-community, langchain-chroma, …).

What changed since the previous edition of these notes

The previous edition pinned langchain==1.3.14, langchain-core==1.5.3, langgraph==1.2.10. All 28 of its examples produce identical output on the versions pinned here. Notable additions in between, none of which change the APIs documented in the three pages:

• langchain 1.3.14 → 1.4.0: the langchain.mcp namespace (MCPAdapter, as_langchain_tool); ContextEditingMiddleware accepts a custom token_counter; standard model exception types.
• langchain-core 1.5.3 → 1.6.2: standard model exception types; StructuredTool is JSON-serializable; assorted fixes to tool-schema generation and content-block handling.
• langgraph 1.2.10 → 1.2.11: trace_policy on add_node; langgraph-checkpoint 4.2.0.