OpenAI AgentKit Release Notes

OpenAI has introduced AgentKit, a comprehensive set of tools designed to help developers and enterprises build, deploy, and optimize AI agents. This release addresses the fragmentation of agent development by replacing complex manual orchestration and custom connectors with a visual design canvas and standardized building blocks.

Agent Design and Orchestration with Agent Builder

Agent Builder is a visual canvas that allows developers to compose agent logic using drag-and-drop nodes, connect tools, and configure custom guardrails. It is designed to accelerate iteration cycles by supporting preview runs, inline evaluation configuration, and full versioning.

According to Ramp, the tool reduced the time required for complex orchestration and manual optimizations from months to a few hours, slashing iteration cycles by 70%. Similarly, LY Corporation reported building their first multi-agentic workflow in less than two hours.

Enterprise Governance and Safety

To support enterprise-scale deployment, OpenAI introduced the Connector Registry, a central admin panel for managing data sources across ChatGPT and the API. This registry includes pre-built connectors for services such as Google Drive, Dropbox, Sharepoint, and Microsoft Teams, as well as third-party Model Context Protocol (MCP) connectors.

For safety, developers can integrate Guardrails, an open-source, modular safety layer. Guardrails can be deployed standalone or via Python and JavaScript libraries to detect jailbreaks, flag personally identifiable information (PII), and prevent unintended behaviors.

Embedding Agentic Experiences with ChatKit

ChatKit is a toolkit for embedding customizable, chat-based agent experiences directly into third-party applications or websites. It handles complex UI requirements such as streaming responses, thread management, and the visualization of model thinking processes.

Canva reported saving over two weeks of development time by using ChatKit to integrate a support agent for its developers community in less than an hour.

Performance Measurement via Enhanced Evals

OpenAI has expanded its Evals platform with four new capabilities to help developers build production-ready agents through rigorous testing:

  • Datasets: Tools to build agent evaluations from scratch and expand them using human annotations and automated graders.
  • Trace Grading: End-to-end assessments of agentic workflows to pinpoint specific failure points through automated grading.
  • Automated Prompt Optimization: The ability to generate improved prompts based on grader outputs and human annotations.
  • Third-party Model Support: The capability to evaluate models from other providers within the OpenAI Evals platform.

Carlyle reported that these evaluation tools increased agent accuracy by 30% and reduced development time for their multi-agent due diligence framework by over 50%.

Model Optimization with Reinforcement Fine-Tuning (RFT)

Reinforcement Fine-Tuning (RFT) allows developers to customize reasoning models. RFT is generally available for OpenAI o4-mini and is in private beta for GPT-5. Two new features have been added to the RFT beta to further enhance agent performance:

  • Custom Tool Calls: Training models to invoke the correct tools at the optimal time to improve reasoning.
  • Custom Graders: The ability to define specific evaluation criteria tailored to a particular use case.

Availability and Pricing

ChatKit and the new Evals capabilities are generally available to all developers. Agent Builder is currently in beta, and the Connector Registry is rolling out in beta to API, ChatGPT Enterprise, and Edu customers who have a Global Admin Console. All AgentKit tools are included under standard API model pricing.

Update (June 3, 2026): OpenAI is winding down Agent Builder and Evals products. These will no longer be available on the OpenAI platform after November 30, 2026. Developers are encouraged to migrate to the Agents SDK for code-based workflows or Workspace Agents in ChatGPT for natural language prompting.

Sources