| name | langchain-architecture |
| description | Design LLM applications using LangChain 1.x and LangGraph for agents, memory, and tool integration. Use when building LangChain applications, implementing AI agents, or creating complex LLM workflows. |
LangChain & LangGraph Architecture
Modern LangChain 1.x and LangGraph for LLM applications with agents, state management, memory, and tool integration.
Core Rules
- LangGraph is the standard for agents: explicit
StateGraph with TypedDict state, durable execution, checkpointing, human-in-the-loop
- Use
create_react_agent for tool-calling agents; define tools with Pydantic schemas
- Memory via checkpointers:
MemorySaver in development, PostgresSaver in production, vector store for long-term context; each thread_id is a separate conversation
- Multi-agent systems use a supervisor node routing between specialized agents via conditional edges
- Document pipelines: loaders, text splitters, vector stores, retrievers
- LangSmith for observability; keys always from env, never inline
References
Read only the file relevant to the task:
| File | Read when working on |
|---|
references/agents-and-state.md | Package structure, TypedDict state, ReAct agent quick start, structured tools with Pydantic |
references/graph-patterns.md | RAG graphs, multi-step workflows with conditional routing, multi-agent supervisor orchestration |
references/memory.md | Checkpointers (MemorySaver, PostgresSaver), thread IDs, vector store long-term memory |
references/observability-performance.md | LangSmith tracing, custom callbacks, streaming, Redis LLM cache, async batching, connection pooling |
