Develop Production-Ready LangChain Agents

This skill provides patterns, checklists, and implementation guidance for LangChain 0.1+ and LangGraph agent development. It includes code examples for state management, RAG pipelines, tool integration, memory systems, and FastAPI deployment patterns. The source material covers ReAct agents using create_react_agent, plan-and-execute patterns with separate planning and execution nodes, multi-agent orchestration with supervisor routing via Command, and various memory types including ConversationTokenBufferMemory, ConversationSummaryMemory, ConversationEntityMemory, and VectorStoreRetrieverMemory.

Use this skill when working on LangChain/LangGraph agent development tasks or workflows, or when needing guidance, best practices, or checklists for LangChain/LangGraph agent development.

Do not use this skill when the task is unrelated to LangChain/LangGraph agent development or when you need a different domain or tool outside this scope.

You provide your agent development goals, constraints, and required inputs as arguments. The skill clarifies goals, constraints, and required inputs, applies relevant best practices and validates outcomes, and provides actionable steps and verification. If detailed examples are required, it references resources/implementation-playbook.md.

LLM and Embeddings: The source material specifies Claude Sonnet 4.5 (claude-sonnet-4-5) via langchain_anthropic.ChatAnthropic as the primary LLM. Voyage AI embeddings are specified with voyage-3-large noted as officially recommended by Anthropic for Claude, plus specialized models voyage-code-3, voyage-finance-2, and voyage-law-2 via langchain_voyageai.VoyageAIEmbeddings.

Vector Stores: Pinecone via langchain_pinecone.PineconeVectorStore with hybrid search support (search_type="hybrid").

Observability: LangSmith for tracing agent executions; Prometheus for tracking metrics (requests, latency, errors); structlog for structured logging.

Deployment: FastAPI with StreamingResponse for streaming agent responses (media_type="text/event-stream").

Caching and Persistence: Redis for response caching with TTL; LangGraph checkpointers for state versioning.

Reranking: Cohere Rerank for relevance optimization in RAG pipelines.

Memory: ConversationTokenBufferMemory for token-based windowing, ConversationSummaryMemory for compressing long histories, ConversationEntityMemory for tracking people, places, and facts, and VectorStoreRetrieverMemory with semantic search.

Here's a state management pattern from the source:

from langgraph.graph import StateGraph, MessagesState, START, END
from langgraph.prebuilt import create_react_agent
from langchain_anthropic import ChatAnthropic

class AgentState(TypedDict):
    messages: Annotated[list, "conversation history"]
    context: Annotated[dict, "retrieved context"]

And a RAG setup example:

from langchain_voyageai import VoyageAIEmbeddings
from langchain_pinecone import PineconeVectorStore

# Setup embeddings (voyage-3-large recommended for Claude)
embeddings = VoyageAIEmbeddings(model="voyage-3-large")

# Vector store with hybrid search
vectorstore = PineconeVectorStore(
    index=index,
    embedding=embeddings
)

# Retriever with reranking
base_retriever = vectorstore.as_retriever(
    search_type="hybrid",
    search_kwargs={"k": 20, "alpha": 0.5}
)

This skill is for users working on LangChain and LangGraph agent development tasks requiring the patterns, integrations, and deployment approaches documented in the source material.