llm-application-dev-langchain-agent

Use create_react_agent from langgraph.prebuilt with an LLM and a list of StructuredTool instances. For custom control flow, define a StateGraph, add nodes and conditional edges, then compile with a checkpointer for state persistence.

Use ConversationSummaryMemory to compress long histories, or ConversationTokenBufferMemory for token-based windowing. For semantic recall, VectorStoreRetrieverMemory retrieves relevant past exchanges via embeddings, and hybrid setups combine multiple types.

Yes, LangChain provides ainvoke and astream for asynchronous invocation and token streaming. Combine these with FastAPI StreamingResponse to serve server-sent events from agent endpoints.

Enable LangSmith tracing to record all agent executions, add Prometheus metrics for latency and errors, and use structlog for structured logging. Health checks should validate the LLM, tools, memory, and external services.

Use multi-agent orchestration when tasks require specialized agents with a supervisor routing between them via Command. A single ReAct agent is better for general-purpose tasks with straightforward tool usage.