OpenAI's airline bot demo shows how agent handoffs actually work

What it does

This is OpenAI’s official demo of a multi-agent customer service system for airlines. A Python backend orchestrates six specialist agents—triage, flight info, booking, seat services, FAQ, and refunds—using the OpenAI Agents SDK. A Next.js frontend visualizes the handoffs and provides a chat interface via ChatKit.

The interesting bit

The demo is deliberately designed to fail well. It includes explicit guardrail triggers—relevance checks that block strawberry poems, jailbreak tests that catch prompt-injection attempts—so you can watch the safety layer trip in real time. The irregular-operations flow (delayed connection, automatic rebooking, compensation) shows how multi-step agent chaining handles edge cases that break single-bot designs.

Key highlights

• Six specialized agents with defined handoff rules, not one generalist model pretending to know everything
• Visual UI shows routing decisions as they happen; useful for debugging why an agent was chosen
• Includes working guardrail examples (relevance + jailbreak) with visible failure states
• Mock data supports two complete scenarios: routine requests and a full disruption/rebooking flow
• Backend can run standalone (FastAPI/uvicorn) if you want to bring your own frontend

Caveats

• Explicitly “designed for demonstration purposes”—all flight data is mocked, not connected to live systems
• Contributing note warns that PRs may not be reviewed, so don’t expect active maintenance

Verdict

Worth cloning if you’re building multi-agent systems and need a concrete reference for handoff logic and guardrail placement. Skip it if you want production airline integrations; the value is in the architecture pattern, not the domain implementation.