OpenAI unveils AgentKit

OpenAI has announced AgentKit, a powerful suite of tools designed to make it easier for developers and enterprises to build, deploy, and optimize AI agents. The new toolkit addresses a key challenge in the AI space—fragmented workflows that make agent development slow and complex.

With AgentKit, developers can now visually design multi-agent workflows, embed conversational UIs, and measure performance—all from a unified platform. It builds on OpenAI’s earlier Responses API and Agents SDK, expanding them into an integrated framework for creating production-ready agents.

What is AgentKit?

AgentKit is a comprehensive developer suite that simplifies how AI agents are built and managed. It combines visual design tools, data connectors, and evaluation features into one ecosystem.

The toolkit introduces three main components—Agent Builder, Connector Registry, and ChatKit—alongside improved evaluation and fine-tuning capabilities. Together, these tools reduce development time and make AI agent deployment more reliable and scalable.

Design workflows visually with Agent Builder

The Agent Builder lets developers design workflows using a drag-and-drop interface. Each node represents a function, logic path, or tool, allowing teams to map out multi-agent processes visually.

The builder supports preview runs, custom guardrails, and full versioning, helping developers iterate faster and maintain control over evolving workflows.

Early adopters like Ramp and LY Corporation have reported dramatic time savings. Projects that once took months now take hours, with better collaboration between product, legal, and engineering teams.

Manage data connections with Connector Registry

Enterprises can use the new Connector Registry to manage how data flows across tools and workspaces. The registry centralizes connectors for services like Google Drive, Dropbox, Microsoft Teams, and SharePoint.

It also supports third-party integrations through Modular Connector Protocols (MCPs), allowing seamless data governance across multiple teams.

This is particularly valuable for large organizations managing multiple AI deployments across departments or subsidiaries, ensuring security and compliance remain intact.

Build chat-based agents with ChatKit

For developers who want to integrate AI chat experiences into apps or websites, ChatKit offers ready-to-use building blocks. It simplifies everything from handling streaming responses to customizing themes.

Companies like Canva, HubSpot, and LegalOn are already using ChatKit to transform static support systems into conversational agents. Canva, for example, integrated a support agent for its developer community in less than an hour using ChatKit.

Evaluate and optimize with new Evals tools

OpenAI has also upgraded its Evals platform to help developers measure and improve agent performance. The new features include Datasets, Trace Grading, Automated Prompt Optimization, and Third-Party Model Support.

These tools make it easier to test prompts, analyze reasoning steps, and refine model outputs for better accuracy. Companies like Carlyle have seen over 30% accuracy improvements and reduced development time by half using these evaluation tools.

Fine-tune reasoning with reinforcement training

Another highlight is Reinforcement Fine-Tuning (RFT), now available on OpenAI o4-mini and in private beta for GPT-5. Developers can customize reasoning models and set custom evaluation metrics tailored to specific use cases.

The beta also introduces Custom Tool Calls and Custom Graders, which improve how agents decide which tools to use and how performance is measured.

Availability and pricing

AgentKit’s tools are being rolled out in phases.

• ChatKit and new Evals capabilities are generally available.
• Agent Builder is currently in beta.
• Connector Registry is launching gradually for API, ChatGPT Enterprise, and Edu users with Global Admin Console access.

All tools are included under standard API pricing. OpenAI also plans to release a standalone Workflows API and new deployment options within ChatGPT soon.