LangChain

In v1 the langchain package was trimmed to agent building blocks, with older chains, retrievers and the indexing API moved to langchain-classic. The central API is create_agent, which returns a LangGraph graph that loops between the model and tools. Middleware hooks into that loop to change prompts, filter tools, summarize history, enforce limits or pause for approval.

Multi-agent systems are built from these agents. The docs list five patterns: subagents (a main agent calls other agents wrapped as tools and keeps control), handoffs (tool calls change state that switches the active agent or its configuration), skills loaded on demand, routers that classify and dispatch, and custom workflows written directly in LangGraph. For a batteries-included option the docs point to Deep Agents, which is built on LangChain.

State comes from LangGraph: a checkpointer gives thread-level short-term memory and makes interrupts possible, and a store provides long-term memory across threads.

HumanInTheLoopMiddleware checks each proposed tool call against an interrupt_on policy. When a call needs review, the middleware raises a LangGraph interrupt and the agent state is saved by the checkpointer. A person then answers each pending action with approve, edit (change the arguments), reject (skip with feedback) or respond (reply in place of an ask-the-user tool), and the run resumes with those decisions in order. A checkpointer and thread ID are required. The docs also cover rendering these approvals in a frontend.

pip install -U langchain

from langchain.agents import create_agent
from langchain.tools import tool

researcher = create_agent(model="openai:gpt-5.5",
                          system_prompt="Answer with three short factual bullet points.")

@tool
def research(question: str) -> str:
    """Ask the research subagent a factual question."""
    result = researcher.invoke({"messages": [{"role": "user", "content": question}]})
    return result["messages"][-1].content

writer = create_agent(model="openai:gpt-5.5", tools=[research],
                      system_prompt="Call research() for facts, then write one paragraph.")

# needs OPENAI_API_KEY and `pip install "langchain[openai]"`
out = writer.invoke({"messages": [{"role": "user", "content": "Explain the A2A protocol."}]})
print(out["messages"][-1].content)

• Requires Python 3.10+; v1 dropped Python 3.9.
• Model providers are separate packages or extras (langchain-openai, langchain[openai]) and need their API key, for example OPENAI_API_KEY.
• Legacy chains, retrievers, the indexing API and langchain-community exports moved to langchain-classic in v1.
• langgraph.prebuilt.create_react_agent is superseded by create_agent; follow the v1 migration guide.
• The built-in langchain.mcp namespace is beta and emits a LangChainBetaWarning; code on langchain-mcp-adapters has a migration guide.
• Human-in-the-loop middleware needs a checkpointer and a thread ID to pause and resume.

Yes. MCPAdapter in langchain.mcp (beta, langchain[mcp]>=1.4.0) loads tools from MCP servers into agents; it replaces the older langchain-mcp-adapters package.

LangChain's create_agent is a ready-made agent loop built on LangGraph, customised with middleware. LangGraph is the lower-level runtime for writing your own graphs.

The library is MIT-licensed. LangChain Inc. sells LangSmith for tracing, evaluation and deployment, with a free developer tier and paid plans.

Yes. LangChain.js is a separate repository (langchain-ai/langchainjs); this record covers the Python package.

Into langchain-classic, which holds legacy chains, retrievers, the indexing API, the hub module and langchain-community exports.