Hands-On Guide to Building and Deploying with OpenAI’s AgentKit

Go beyond single-prompt chatbots. This guide unlocks visual orchestration, detailing how to build, test, and deploy complex, multi-agent AI systems with newfound ease.

OpenAI’s AgentKit marks a major leap in how developers can design and deploy AI agents. Earlier, the AgentSDK required writing detailed orchestration logic and manually wiring tools. With AgentKit, much of that complexity is replaced by a visual drag-and-drop interface that feels almost effortless. In this hands-on guide, we will create a simple workflow to see how easily agents can be built, tested, and deployed. We will also explore how ChatKit brings customizable chat experiences directly into applications.

Table of Contents

  • The Journey to AgentKit
  • What Makes AgentKit Different
  • Setting Up Your Environment
  • Building an Assistant for Developers
  • Exploring ChatKit
  • Testing and Deploying Options

The Journey to AgentKit

OpenAI’s journey toward AgentKit began with simple language models that could only generate text responses. As developers sought more control and interactivity, OpenAI introduced APIs and tools that allowed models to call functions and use external data. The release of the AgentSDK in early 2025 was a turning point because it enabled developers to orchestrate multiple agents and define complex workflows. However, it still required significant coding effort. AgentKit builds on this foundation by adding a visual workflow builder, integrated evaluation tools, and ChatKit for seamless deployment, making the entire agent development process more intuitive and accessible.

Evolution from GPT models to AgentKit

Evolution from GPT models to AgentKit

What Makes AgentKit Different

AgentKit distinguishes itself by offering a more integrated and accessible platform approach to building AI agents. Unlike code-intensive frameworks such as Autogen or crewAI, which excel at orchestrating complex multi-agent conversations and collaborations, AgentKit aims to simplify the entire lifecycle from creation to deployment. While LlamaIndex is primarily a data framework for RAG, and LangGraph provides granular control over stateful flows, AgentKit abstracts much of this complexity. It focuses on a tool-centric, goal-oriented model that empowers even non-developers. This makes it fundamentally different from workflow automation tools like n8n, as AgentKit is purpose-built for creating specialized, human-like AI agents, not just connecting APIs.

Setting Up Your Environment

To start using AgentKit, you first need an OpenAI developer account, which can be created at platform.openai.com. Ensure you have API access and a modern web browser for the visual workflow interface. The prerequisites include a basic understanding of AI agents and a working internet connection. AgentKit is available under standard OpenAI API pricing with no extra charges for the platform itself. Once logged in, you can access the AgentKit dashboard to create, test, and manage agents in a secure and user-friendly environment.

Builder page layout

Builder page layout

Building an Assistant for Developers

To demonstrate a practical application, we will design an intelligent query resolution workflow for a technology company using AgentKit. This system must autonomously differentiate between two primary inquiry streams, one being Human Resources (HR) requests concerning leave and payslips, and the other being technical support questions related to data architecture or bug resolution. The core objective is to evaluate AgentKit’s efficacy and ease of implementation in creating a robust workflow where specialized agents come together to provide a seamless user experience.

Setting Up the Workflow

A sample workflow

A sample workflow

Our query resolution workflow initiates with a primary agent that receives the user’s request. This request is immediately passed to a classifier agent, which analyzes the content and categorizes it as either ‘HR’ or ‘Tech’. Based on this classification, a conditional routing node directs the query down the appropriate path. ‘HR’ inquiries are funneled to a dedicated HR Agent. At the same time, technical questions are handled by the Tech Agent, ensuring that users receive a final, expert response from the correct source.

Considering the Data Types

When configuring each agent, it’s crucial to consider its data output, as this becomes the direct input for subsequent nodes in the workflow. You can select either text for simple, direct string-based outputs or JSON for structured data, which is ideal for passing multiple fields reliably. Additionally, there is a natural language support, even for creating JSON schemas.

Setting up JSON schema

Setting up JSON schema

For this demonstration, we’ve set the classifier agent’s output to text. We strictly constrain this output to be only ‘HR’ or ‘Tech’. Consequently, the next if/else node can easily route the query using a simple conditional check on that text.

Choosing the Right Model

Selecting the right model is crucial for optimizing your workflow’s performance. For complex, in-depth research tasks, leveraging a powerful, state-of-the-art model is most effective. For general-purpose or miscellaneous tasks, a versatile model like a future GPT iteration would be a strong choice. However, for smaller, well-defined jobs, such as classifying a query into ‘HR’ or ‘Tech’, a compact and efficient model like ‘GPT-4o-mini’ can be ideal. This not only ensures faster performance but also leads to significant cost optimization.

Configuring instructions and model

Configuring instructions and model

Exploring the Tools

Agents derive their power from the tools they can access. AgentKit provides a versatile toolkit, including hosted tools managed by OpenAI like web-search, file-search, and code-interpreter, which offer powerful, ready-to-use capabilities. It also supports local tools, such as function or custom, allowing you to integrate your own specific logic or APIs.

Configuring instructions and model

Exploring the tool options

For our use case, the Tech Agent can be customized to file-search internal wikis and use a custom tool to query a bug-tracking system. Similarly, the HR Agent can be equipped with file-search to exclusively access uploaded company policy documents, ensuring its responses on leave or payroll are always accurate and source-based.

Testing the Flow of Information

To ensure the workflow operates as intended, you can use the Preview feature to conduct end-to-end testing. Clicking “Preview” opens a chat interface where you can input a sample query. You can then visually trace the flow of information in real-time as it moves through each node, from classification to conditional routing and final processing, before receiving the generated output.

Testing the workflow with a query

Testing the workflow with a query

Note that this feature may initially be disabled, requiring organization verification. To resolve this, simply hover over the “Preview” button to begin the process. This one-time verification requires submitting a valid government-issued ID and completing a face scan.

Exploring ChatKit

ChatKit is the final step in deploying your agent, allowing you to create a user-facing chatbot interface. To begin, you must add the Client tool to your agent’s configuration; this tool is essential for connecting your backend logic to the frontend chat widget.

Testing the workflow with a query

Exploring the deployment options

Once enabled, clicking the ‘Code’ tab in the ChatKit section provides an embeddable HTML snippet for your website. For security, you must use the ‘Add an allowed domain’ feature. This requires you to whitelist the specific domains where the chatbot can operate, ensuring it only functions on your authorized websites and preventing unauthorized embedding elsewhere.

Testing and Deploying Options

AgentKit streamlines the transition from testing to deployment. For developers who prefer hands-on control, the platform can generate the complete workflow as code, identical to what you would write manually using the AgentSDK. This allows for deep integration into existing codebases.

Alternatively, for a much simpler deployment, AgentKit provides a unique ‘workflowID’. Using this ID, you can directly integrate and run the entire workflow in your application with just a few lines of code, abstracting away the underlying complexity and enabling rapid production deployment.

Way Forward

This guide is your launchpad into the world of intuitive AI development. With AgentKit, you are now equipped to build and deploy sophisticated multi-agent systems that solve real-world problems, from internal tools to customer-facing chatbots. This is just the beginning. The future scope includes creating fully autonomous agents that manage complex business processes and proactive assistants that anticipate user needs. The era of visually orchestrated AI is here. Embrace the tools and start building the intelligent solutions of tomorrow.

References

Picture of Abhishek Kumar

Abhishek Kumar

Abhishek is an AI and analytics professional with deep expertise in machine learning and data science. With a background in EdTech, he transitioned from Physics education to AI, self-learning Python and ML. As Manager cum Assistant Professor at Miles Education and Manager - AI Research at AIM, he focuses on AI applications, data science, and analytics, driving innovation in education and technology.

The Chartered Data Scientist Designation

Achieve the highest distinction in the data science profession.

Elevate Your Team's AI Skills with our Proven Training Programs

Strengthen Critical AI Skills with Trusted Generative AI Training by Association of Data Scientists.

Our Accreditations

Get global recognition for AI skills

Chartered Data Scientist (CDS™)

The highest distinction in the data science profession. Not just earn a charter, but use it as a designation.

Certified Data Scientist - Associate Level

Global recognition of data science skills at the beginner level.

Certified Generative AI Engineer

An upskilling-linked certification initiative designed to recognize talent in generative AI and large language models

Join thousands of members and receive all benefits.

Become Our Member

We offer both Individual & Institutional Membership.

ADaSci Certified Agentic AI System Architect

The ADaSci Certified Agentic AI System Architect program is a 30-hour, self-paced certification designed to equip professionals with the skills to design, deploy, and manage scalable agentic AI systems.