Multi-agent orchestration with LangGraph: the design

What LangGraph is

LangGraph models an agent as a state graph: nodes that do work, edges that decide the next step and a shared state that travels between them. With a checkpointer, that state can be saved and resumed, which is what gives a conversation memory.

How I designed it for the platform

In the AI agents platform for companies, the agent service is a LangGraph state graph. An orchestrator receives the question and routes it to one of three agents: SQL, reports (charts and PDF) or RAG. Memory uses a checkpointer.

Around the graph there is an input guardrail against prompt injection and an output guardrail that validates and cites, and traces and costs go to Langfuse.

What is done and what is not

Nux, the Bonuxo assistant, also has the LangGraph integration in progress; it is not finished and its LLM layer is switched off in production.