What is OpenAI's AgentKit, and how do you build an agent with Agent Builder?

AgentKit is a set of OpenAI tools, launched on October 6, 2025, for designing, deploying and improving AI agents; its main pieces are Agent Builder, a visual canvas for agent workflows, ChatKit, a chat window you can embed in your own product, a Connector Registry for managing data access, and new evaluation tools. It builds on OpenAI's Responses API and is included in standard API pricing.

What is OpenAI's AgentKit, and how do you build an agent with…

Yesterday, Monday, October 6, 2025, at its DevDay conference in San Francisco, OpenAI launched AgentKit, a toolkit for companies and developers who want to build AI agents. An agent is an AI system that works toward a goal over several steps, deciding which tools to use, such as searching documents or looking up an order, instead of giving a single reply.

OpenAI's pitch is that building agents has meant stitching together many separate pieces: code to coordinate the steps, custom connections to data, hand-built testing, endless prompt tweaking and weeks of interface work. AgentKit puts those pieces in one place. Sam Altman described it on stage as a full set of building blocks to take agents from prototype to production.

The pieces of AgentKit

Piece	What it does	Status
Agent Builder	A visual canvas where you drag and connect steps, tools and safety checks, test runs, and keep versions	Beta
ChatKit	A ready-made chat interface you embed in your app or website and style to match your brand	Available to all developers
Connector Registry	One admin panel for the data sources and tools agents may use, such as Google Drive, Dropbox, SharePoint, Microsoft Teams and MCP servers	Starting a limited beta for some business customers
Guardrails	An open-source safety layer that can hide or flag personal information and detect jailbreak attempts	Usable in Agent Builder or as a Python or JavaScript library
Evals	Tools to measure and improve an agent: datasets, trace grading, automatic prompt optimization and support for other companies' models	Available to all developers

MCP, the Model Context Protocol, is an open standard for plugging tools and data into AI systems. A jailbreak is an attempt to trick a model into ignoring its rules. All of this sits on the Responses API, the developer interface OpenAI released in March together with its Agents SDK.

How Agent Builder works

Altman compared Agent Builder to Canva, the design tool: instead of writing code, you lay out the agent's logic visually. Each box on the canvas, called a node, is one step, such as an AI model call, a tool, a decision or a guardrail. Arrows show the order in which they run. You can start from a blank canvas or a template, run a preview to watch each step, attach tests directly to the workflow, and save versions so you can roll back.

OpenAI's examples are about speed. Ramp says it went from a blank canvas to a working purchasing agent in a few hours, and LY Corporation in Japan built a work assistant in under two hours. On stage, OpenAI's Christina Huang built a workflow with two agents in under eight minutes.

A worked example: a support agent for a small online shop

Here is how a builder might lay out a customer-support agent in Agent Builder. Each step is a node on the canvas:

Start: receives the customer's message from the chat window.
Guardrail: masks personal details like card numbers and blocks jailbreak attempts before anything reaches the model.
Classifier agent: decides whether the message is about an order, a return or something else.
Decision: sends order questions to the order agent, returns to the returns agent, and everything else to a human.
Order agent: uses a connector to look up the order status, then answers.
Returns agent: reads the shop's returns policy file and drafts a reply, following instructions like these:

Prompt

You handle return requests for an online bookshop.
Use only the returns policy provided. If the item is outside the
30-day window, say so politely and offer store credit instead.
Never promise a refund for damaged items; pass those to a human.
Answer in three sentences or fewer.

Publish: the finished workflow is placed on the shop's website through ChatKit, so customers chat with it in a window that looks like the rest of the site.

Before going live, the builder creates a small dataset, say twenty real past questions with good answers. Trace grading then scores each full run, step by step, so you can see whether a wrong answer came from the classifier or the returns agent. Automatic prompt optimization suggests better instructions based on those grades and on people's comments.

Try it: Sketch the support agent above on paper for a business you know. Draw one box per step, mark where a person must take over, and write five test questions with the answers you would expect.

Why it matters

TechCrunch frames AgentKit as part of a race among AI companies to offer complete agent-building platforms to businesses, so that agents do real tasks rather than just chat. For OpenAI, it is a way to keep developers on its platform; Altman said it is everything the company wished it had when it built its first agents.

For builders, the main change is who can take part. A visual canvas lets product managers, legal reviewers and engineers look at the same workflow. Ramp says this kept its teams aligned and cut the time for each round of changes by 70%.

There are limits to keep in mind. Agent Builder is a beta, the Connector Registry requires OpenAI's new Global Admin Console, and OpenAI says a standalone Workflows API and ways to deploy agents into ChatGPT are still to come. As with any agent, the quality of your tests decides how far you can trust it.

How do you start building your first agent with AgentKit?

Sign in to the OpenAI platform with a developer account and open Agent Builder. Begin with a template close to your goal rather than a blank canvas, and pick one narrow job, such as answering questions from a single policy document.

Add a guardrail early, run previews, and watch every step until the behavior is predictable. Then build a small evaluation dataset from real examples and use trace grading to find the weak step before you change anything.

When the results are steady, embed the agent with ChatKit on a test page, let a few colleagues try it, and only then put it in front of customers. Keep a clear route to a human for anything the agent should not decide alone. AgentKit tools are charged at standard API model prices, so your cost depends on how many model calls the workflow makes.