# AgentKit Demo

Source: https://www.youtube.com/watch?v=DwGNK1DUFsM
Recap page: https://rapidrecap.app/video/DwGNK1DUFsM
Generated: 2025-10-08T18:03:31.667+00:00

---
## Quick Overview

Christina Huang demonstrates building a fully functional AI agent workflow using AgentKit in under eight minutes, which involves chaining specialized agents for session retrieval and general information, integrating guardrails for safety, and customizing the final output as a visual widget, all deployable without manual code changes.

**Key Points:**
- Christina Huang builds a complete AI agent workflow in under 8 minutes using AgentKit during the OpenAI DevDay 2025 presentation.
- The workflow utilizes two specialized agents: a 'Session Agent' for retrieving specific session data and a 'DevDay Agent' for general information, styled to sound like the character Froge.
- A 'Guardrails' node is integrated to protect against hallucinations, PII exposure, and jailbreaking, ensuring safe agent behavior.
- The output of the second agent is visually customized using the 'Widget Builder' tool, which allows for aesthetic control over the final chat interface.
- The final workflow is tested live, showing the agent correctly categorizing a query about building agents and retrieving relevant session details, all within the 8-minute constraint.
- The completed workflow is published directly from the visual builder, providing a deployable artifact without requiring external code edits.

![Screenshot at 00:15: Christina Huang introduces the goal of building and shipping an agent using AgentKit in under eight minutes during her presentation.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-00-15.png)

**Context:** The video captures a live demonstration by Christina Huang from OpenAI's Platform Experience team during OpenAI DevDay [2025]. The focus is showcasing the rapid development capabilities of AgentKit, a visual workflow builder designed to help developers quickly create and deploy sophisticated, multi-step AI agents.

## Detailed Analysis

Christina Huang walks through creating an AI agent workflow in AgentKit, starting with the goal of answering attendee questions about the DevDay schedule. She first uses a 'Categorize' agent to route incoming messages based on whether they ask about a specific session or general information. For session-specific queries, the flow routes to the 'Session Agent,' which is configured to use a File Search tool to query a document containing the entire day's schedule. For general queries, the flow routes to the 'DevDay Agent.' This second agent is instructed to answer questions based only on the provided DevDay information and to speak in the persona of 'Froge.' Crucially, a 'Guardrails' node is added before categorization to block sensitive PII and prevent hallucinations or jailbreaks, ensuring safe responses. Finally, the output from the 'DevDay Agent' is customized using the Widget Builder, allowing for visual styling (like a 'Froge' theme) and presenting the final output as an interactive widget directly in the UI. The entire process, including testing the published workflow on a sample website, is completed within the 8-minute time limit.

### Agent Workflow Setup

- Initial categorization node directs queries to specialized agents
- Session Agent uses File Search tool for schedule data
- DevDay Agent handles general queries with a 'Froge' persona.

### Implementing Safety and Logic

- Guardrails node added to block PII, hallucinations, and jailbreaking attempts
- An 'If/Else' node routes traffic based on the categorization output ('session_info' or else).

### Customizing Output with Widget Builder

- The final agent output is configured to render as a visual widget
- Custom styling applied including colors, typography ('OPM_SKE_FONTS'), and placeholder text ('Rabbit robot').

### Deployment and Testing

- The completed workflow is published directly from the visual interface
- Live testing on a sample website confirms the agent correctly answers a query about sessions on building agents, adhering to the 'Froge' persona.

![Screenshot at 00:08: Christina Huang taking the stage at the OpenAI DevDay event before a large audience.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-00-08.png)
![Screenshot at 00:15: Christina Huang stating the challenge: building and shipping an agent in 8 minutes using AgentKit.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-00-15.png)
![Screenshot at 00:34: The initial workflow structure showing the Start node connected to a Categorize agent and an If/Else node.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-00-34.png)
![Screenshot at 01:13: The workflow canvas showing the initial structure with the Categorize agent output branching to an If/Else node.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-01-13.png)
![Screenshot at 01:54: Configuring the If/Else node to branch based on the 'session\_info' output from the Categorize agent.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-01-54.png)
![Screenshot at 02:06: Adding the 'Session Agent' node to handle the 'session\_info' branch of the logic.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-02-06.png)
![Screenshot at 02:23: The 'Session Agent' configured with the 'Sessions' tool attached, containing the DevDay schedule file.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-02-23.png)
![Screenshot at 04:24: The Guardrails node configuration panel showing the 'Personally identifiable information \(PII\)' check enabled.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-04-24.png)
![Screenshot at 06:40: Viewing the generated code export option, showing TypeScript imports for various components.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-06-40.png)
![Screenshot at 08:06: The final successful output displayed in the preview chat window when asking about sessions.](https://ss.rapidrecap.app/screens/DwGNK1DUFsM/00-08-06.png)