In this post we explore the “agent-as-a-tool” pattern using the OpenAI Agents SDK. The pattern treats specialist agents as reusable tools that a manager agent can call, allowing clear division of responsibility when tasks are well‑defined but their internal steps are not.

First we build a Travel planner agent that creates a layover itinerary. The scenario is a family with a 10-hour layover in Munich, needing to leave the airport, sightsee, and eat without missing the flight. To answer feasibility, activity options, and risk questions we define three specialist agents.

The logistics specialist checks transport timing with a web search; the local experience specialist suggests activities and food; the risk specialist reviews the plan for practical issues. Each specialist is exposed to the planner via as_tool, so the planner continues to call a tool and receive a response while the underlying agent performs the work.

This approach gives the travel planner full flexibility without hard‑coding functions, and the system can produce a structured Layover Plan output. The pattern is useful whenever the boundary of a delegated task is clear but the steps to complete it are open‑ended.