I built a react-admin support chatbot with the OpenAI Agents SDK. Here is what each piece looks like in real code, from agents and handoffs to guardrails and tools.

marmelab

A walkthrough of building a multi-agent support chatbot for react-admin using the OpenAI Agents SDK. The architecture includes a TriageAgent that classifies incoming questions and hands off to either a DocsAgent (which searches a vector store of 300+ markdown docs using fileSearchTool) or a SupportAgent for billing/subscription queries. Input and output guardrails are implemented to block prompt injections and prevent information leaks, including OpenAI's omni-moderation model for harmful content filtering. The bot is exposed via a Hono API with a custom React frontend, streaming responses to the UI. The post also touches on tracing, human-in-the-loop, and voice agent features available in the SDK.

A Multi-Agent Support Bot with the OpenAI Agents SDK

<p>Routing through a TriageAgent before the DocsAgent or SupportAgent is a clean way to keep each agent’s context narrow, and putting guardrails on both the input and the output sides is the part people usually skip. Misclassification at triage is where I would expect the quiet failures, since a billing question sent to the docs vector store just returns confident nonsense. How do you handle low-confidence triage, do you fall back to a clarifying question or route to a default agent?</p>
