OpenAI’s Agent Builder is a transformative visual tool that allows developers and creators to build complex AI workflows without writing extensive code. Whether you are starting from a template or building from scratch, this tool provides an all-in-one space to design, test, and launch AI agents visually.
In this tutorial, we will walk through the process of building a sophisticated "Travel Agent" workflow. This system will intelligently classify user requests and route them to specialized agents—one for planning itineraries and another for checking real-time flight data with rich visual outputs.
Setting the Stage: The Classifier Agent
Every great workflow starts with a plan. After launching a new workflow, you begin with a standard "Start" node. The first step in our logic is to determine what the user actually wants. To do this, we create a "Classifier" agent. This agent acts as the traffic controller, analyzing the user's input to decide the next steps.
We configure this Classifier agent with specific instructions to identify if a message is about an "itinerary" or a "flight." To make this machine-readable, we don't just ask for text; we define a structured output using JSON. By creating a property called classification with strict enum options (flight_info or itinerary), we ensure the system makes a definitive choice for every user query.
Building Logic and Branching
Once the Classifier determines the user's intent, the workflow needs to act on it. We utilize an If/Else node to create a logic branch. This node checks the output from our previous step. If the classification equals flight_info, the conversation is routed down one path; if not, it defaults to the itinerary path. This visual branching is powerful because it allows for specialized handling of different tasks within a single conversation interface.
Configuring Specialized Agents
Now that we have our logic branches, we need agents to handle the work. On the itinerary branch, we create a simple agent instructed to build concise travel plans.
On the flight branch, we need more power. We create a "Flight Agent" and, crucially, we equip it with tools. Since flight data changes constantly, a static model isn't enough. We enable the "Web Search" tool, giving this agent the ability to browse the internet for the most up-to-date arrival and departure information.
Testing the Itinerary Workflow
With our structure in place, it is time to test. Using the built-in "Run Preview" feature, we can simulate a user interaction. When we ask for a day plan in Tokyo, we can visually watch the signal travel from the Start node, through the Classifier, down the "Else" branch, and finally to the Itinerary Agent, which returns a text-based travel plan.
Enhancing the Experience with Widgets
Text responses are useful, but for data like flights, a visual card is much better. Agent Builder allows you to import custom widgets to render data beautifully. After designing a flight card in the Widget Studio, we can upload it directly to our Flight Agent node as the output format.
To make the most of this widget, we update the Flight Agent's instructions. We can prompt the AI to be creative with the widget parameters—for example, instructing it to choose a background color for the flight card based on the destination city. This blends the logic of code with the creative reasoning of the LLM.
The Final Result
Now, when we test the workflow with a flight request (e.g., "SFO to Tokyo"), the experience is completely different. The Classifier correctly routes the request, the Flight Agent uses Web Search to find real data, and then it populates our custom widget. The result is a rich, interactive UI element displayed directly in the chat, complete with dynamic styling chosen by the AI.
Deployment
Once satisfied with the workflow, deploying is straightforward. Clicking "Publish" creates a versioned instance of your agent. You are provided with a Workflow ID that you can use to integrate your new AI assistant directly into your product using ChatKit or the Agents SDK, bridging the gap between visual design and production application.
By following these steps, you move beyond simple chatbots to creating intelligent, multi-step workflows that can handle complex tasks and render custom user interfaces.
Top comments (0)