Develop Production-Ready LangChain Agents

This skill builds production-grade AI agent systems with LangChain 0.1+ and LangGraph. Its essential architecture centers on LangGraph state management via a StateGraph/MessagesState-based AgentState TypedDict tracking conversation history and retrieved context, recommending Claude Sonnet 4.5 as the primary LLM and Voyage AI's voyage-3-large embeddings (with domain variants voyage-code-3, voyage-finance-2, voyage-law-2). It covers three agent types - ReAct agents via create_react_agent(llm, tools, state_modifier) for general-purpose multi-step reasoning, Plan-and-Execute with separate planning and execution nodes tracked through state, and Multi-Agent Orchestration using Command[Literal["agent1", "agent2", END]] routing with a supervisor deciding the next agent - and five memory systems (token-windowed ConversationTokenBufferMemory, ConversationSummaryMemory for compressing long histories, ConversationEntityMemory for tracking people/places/facts, VectorStoreRetrieverMemory for semantic search, and hybrid combinations). Its RAG pipeline wires Voyage AI embeddings into a Pinecone vector store with hybrid search and reranking, plus advanced patterns HyDE (hypothetical document generation), RAG Fusion (multiple query perspectives), and Cohere Rerank. It defines tools via StructuredTool.from_function with a Pydantic input schema and async error handling, a FastAPI production server with streaming responses, monitoring via LangSmith tracing, Prometheus metrics, structlog, and health checks, optimization strategies (Redis response caching, vector-DB connection pooling, round-robin load balancing across agent workers, async timeouts, exponential-backoff retry via tenacity), and LangSmith-based evaluation (qa/context_qa/cot_qa evaluators). It closes with a twelve-item implementation checklist and eight best practices, including always using async methods (ainvoke, astream), securing secrets via environment variables, and versioning state with checkpointers for reproducibility.

Use it when building sophisticated, production-grade LangChain/LangGraph agent systems needing async patterns, observability, scalability, and cost optimization. It does not apply outside LangChain/LangGraph agent development - a different domain or tool is out of scope.

Inputs: the target agent's task scope, required tools, and memory/RAG needs. Outputs: a compiled LangGraph StateGraph agent with checkpointing, async tool implementations with Pydantic schemas and error handling, a RAG pipeline with hybrid retrieval and reranking, a FastAPI streaming endpoint, LangSmith tracing and Prometheus metrics, and a LangSmith evaluation suite.

LangChain 0.1+, LangGraph (StateGraph, MessagesState, create_react_agent, checkpointers), langchain_anthropic (Claude Sonnet 4.5), langchain_voyageai embeddings, langchain_pinecone vector store, Cohere Rerank, LangSmith for tracing and evaluation, FastAPI, Redis, Prometheus, structlog, and tenacity for retry logic.

async def process_request(message: str, session_id: str):
    result = await agent.ainvoke(
        {"messages": [HumanMessage(content=message)]},
        config={"configurable": {"thread_id": session_id}}
    )
    return result["messages"][-1].content

Engineers building production LangChain/LangGraph agent systems who need a proven state-graph architecture, memory and RAG design, and a deployment/observability checklist rather than assembling one from scratch.