langchain-architecture

Use create_react_agent from langgraph.prebuilt with an LLM like ChatAnthropic and a list of tools defined via the @tool decorator. Pass a checkpointer such as MemorySaver to enable persistent memory across invocations using thread IDs.

Attach a checkpointer to the agent: MemorySaver for development or PostgresSaver for production. Each invocation uses a configurable thread_id, so conversations with the same thread ID retain full message history.

LangGraph provides explicit typed state, durable execution, human-in-the-loop inspection, and checkpointing, making it the standard for agents. Plain chains suit simple linear prompt-to-response flows without branching or state.

Yes, set streaming=True on the chat model and iterate with astream for token chunks, or use astream_events on agents to receive granular events like on_chat_model_stream and on_tool_start during execution.

Set LANGCHAIN_TRACING_V2=true along with your LANGCHAIN_API_KEY and project name as environment variables. All LangChain and LangGraph operations are then automatically traced with token usage, latency, and error tracking.

Avoid multi-agent supervisor patterns when a single agent with well-scoped tools can complete the task, since routing between agents adds latency and complexity. Reserve it for genuinely distinct specializations like research, writing, and review.