Langchain Knowledge Patch

LangChain 1.0 Knowledge Patch

Claude's baseline knowledge covers LangChain 0.2.x and LangGraph 0.1.x. This skill provides changes from the LangChain 1.0 / LangGraph 1.0 release (2025-10-22).

Quick Reference

Agent Creation

create_agent — Essential Pattern

from langchain.agents import create_agent

agent = create_agent(
    model="anthropic:claude-sonnet-4-6",  # or model instance
    tools=[my_tool],
    system_prompt="You are a helpful assistant",
    middleware=[...],           # replaces pre/post hooks
    response_format=...,       # ToolStrategy or ProviderStrategy
    state_schema=CustomState,  # must be TypedDict, not Pydantic
    context_schema=Context,    # typed runtime context
    name="my_agent",
)

result = agent.invoke(
    {"messages": [{"role": "user", "content": "Hello"}]},
    context=Context(user_id="123")  # replaces config["configurable"]
)

See references/agent-creation.md for full migration guide.

Middleware System

Built-in: SummarizationMiddleware, HumanInTheLoopMiddleware, ModelCallLimitMiddleware, ToolCallLimitMiddleware, ModelFallbackMiddleware, PIIMiddleware, TodoListMiddleware, LLMToolSelectorMiddleware, ToolRetryMiddleware, ContextEditingMiddleware.

All middleware imports from langchain.agents.middleware. See references/middleware.md for decorator and class patterns.

Middleware — Quick Example

from langchain.agents.middleware import before_model, wrap_tool_call

@before_model
def trim_context(state, runtime):
    return {"messages": state["messages"][-20:]}

@wrap_tool_call
def handle_errors(request, handler):
    try:
        return handler(request)
    except Exception as e:
        return ToolMessage(content=f"Error: {e}", tool_call_id=request.tool_call["id"])

agent = create_agent(
    model="claude-sonnet-4-6",
    tools=tools,
    middleware=[trim_context, handle_errors],
)

Tools & Runtime

ToolRuntime gives tools access to agent state, context, and store:

from langchain.tools import tool, ToolRuntime

@tool
def greet(runtime: ToolRuntime[Context, CustomState]) -> str:
    """Greet the user."""
    name = runtime.state.get("user_name", "Unknown")
    return f"Hello {name} (id: {runtime.context.user_id})!"

Reserved parameter names: config, runtime cannot be used as tool argument names.

See references/tools-and-runtime.md.

Content Blocks & Structured Output

Structured Output — Quick Example

from langchain.agents.structured_output import ToolStrategy, ProviderStrategy

# ToolStrategy: works with any model supporting tool calling
agent = create_agent(model="gpt-4.1", tools=tools, response_format=ToolStrategy(MySchema))

# ProviderStrategy: uses native structured output (more reliable, fewer providers)
agent = create_agent(model="gpt-4.1", response_format=ProviderStrategy(MySchema))

# Passing schema directly defaults to ProviderStrategy with ToolStrategy fallback
agent = create_agent(model="gpt-4.1", response_format=MySchema)

result["structured_response"]  # access the typed output

Prompted output is removed. Structured output now happens in the main agent loop (no extra LLM call).

See references/content-and-output.md for content blocks and multimodal messages.

Namespace Changes

langchain-classic

Legacy chains, retrievers, indexing, hub, and community re-exports:

# pip install langchain-classic
from langchain_classic.chains import LLMChain
from langchain_classic.retrievers import ...
from langchain_classic import hub

See references/namespace-and-migration.md for full details.

Breaking Changes (1.0)

• Python 3.10+ required (3.9 dropped)
• message.text is now a property — use response.text not response.text()
• example parameter removed from AIMessage
• langchain-anthropic max_tokens default increased (was 1024)
• Chat model return type fixed to AIMessage (was BaseMessage)
• OpenAI Responses API now defaults to storing response items in message content

Deep Agents SDK

New deepagents package for complex multi-step tasks:

from deepagents import create_deep_agent

agent = create_deep_agent(
    tools=[my_tool],
    system_prompt="You are a helpful assistant",
)

Built-in capabilities: write_todos (planning), ls/read_file/write_file/edit_file (context management), task (subagent spawning). Pluggable filesystem backends (in-memory, local disk, LangGraph store, sandboxes).

See references/namespace-and-migration.md for package details.

Reference Files

Source: Nevaberry/nevaberry-plugins — distributed by TomeVault.