OpenAI introduced AgentKit in October 2025 as a collection of tools for building, deploying, and improving AI agents. The original package included Agent Builder, Connector Registry, ChatKit, expanded evaluation tools, and integrations with the Responses API and Agents SDK.

The idea was to give developers more of the necessary infrastructure in one place. Instead of building orchestration, connectors, evaluations, and chat interfaces from scratch, teams could use OpenAI’s own components.

But the product has changed considerably since launch. On June 3, 2026, OpenAI announced that Agent Builder and its Evals products were being wound down. Both are scheduled to leave the OpenAI platform after November 30, 2026. For workflows that need to continue in code, OpenAI now recommends the Agents SDK.

That changes the picture. AgentKit can no longer be judged solely by what appeared in the original announcement.

For new projects, the OpenAI Agents SDK is now the most relevant part of the stack.

It provides the runtime for agents that plan tasks, call tools, retain state, and work with other specialized agents. Developers get built-in support for repeated tool calls, agent handoffs, sessions, tracing, guardrails, and approval steps.

Consider a customer-support agent. It might investigate a problem, pass the case to a specialist agent, check an internal system, ask a person to approve an action, and then pick up where it left off. That is a workflow, not a single prompt and response.

OpenAI draws a useful distinction between the Agents SDK and the Responses API. With the Responses API, developers control the loop and decide how the workflow branches. The Agents SDK handles more of that recurring orchestration automatically.

Agent Builder was originally one of AgentKit’s headline features. It gave developers a visual canvas for arranging workflow nodes, connecting tools, configuring guardrails, previewing runs, and managing versions of multi-agent systems.

For many teams, that visual approach was easier to inspect than a web of prompts and functions spread across a codebase. You could see the workflow rather than reconstructing it in your head.

Its problem is longevity. OpenAI’s developer documentation now lists Agent Builder under legacy APIs, and the product is scheduled to shut down on November 30, 2026.

Existing users therefore need to plan a migration. For new users, building a long-term project around Agent Builder would mean starting with a tool that already has an end date.

ChatKit is not being retired alongside Agent Builder. OpenAI says it will remain available after Agent Builder is deprecated.

It provides an embeddable interface for agent applications, with support for chat widgets, file attachments, tool calls, customizable prompts, and other interface components. Developers can connect it to their own server-side agent implementation, including one built with the Agents SDK.

The separation makes sense. Agent logic and chat-interface design are different problems. ChatKit handles the familiar conversation layer so teams do not have to rebuild common chat features every time they create an agent application.

The original launch also introduced modular Guardrails for jobs such as detecting jailbreak attempts and masking or flagging personally identifiable information.

The parts of AgentKit that remain relevant suit customer-support agents, internal knowledge assistants, transactional workflows, research agents, approval-based processes, and applications in which several specialized agents need to work together.

The Agents SDK is particularly helpful when a workflow involves repeated tool use, branching, and handoffs. ChatKit is a natural companion when the finished agent also needs a polished conversational interface.

Teams looking for a mostly visual, low-code builder will find the current direction less appealing than the original AgentKit pitch.

The biggest issue is the product transition. A review based only on the 2025 launch would now be misleading because Agent Builder and the Evals platform product are approaching retirement.

The move toward code gives developers more control, but it also raises the entry barrier. Teams attracted to AgentKit because of its visual workflow builder will need to recreate that work with the Agents SDK or choose a different platform.

There is also the underlying problem of agent reliability. Tracing, guardrails, approvals, and evaluation tools help, but they do not make an agent automatically safe. Any workflow that can reach business systems or take actions outside the chat still needs careful testing and tightly designed permissions.