OpenAI AgentKit: Shutdown Dates and Migration Guide

OpenAI AgentKit was OpenAI's collection of tools for building, deploying, and evaluating agents; its Agent Builder service is scheduled to shut down on November 30, 2026. The Agents SDK and ChatKit remain available, but a ChatKit app backed by Agent Builder needs your own replacement backend. Existing evals become read-only on October 31, before the Evals dashboard/API and the v1/prompts API with reusable prompt objects are scheduled to shut down on November 30.

Verified against OpenAI's live documentation on October 14, 2026. The practical job is to preserve your workflow, its prompts, and the tests that prove it still works. Treat those as separate migration tasks with separate owners. OpenAI's deprecation schedule supplies the dates; the checklist below turns them into delivery milestones.

OpenAI AgentKit: What Changed and What Survives

The hosted builder is closing; OpenAI's agent development stack continues. AgentKit grouped tools for building, deploying, and evaluating agents. OpenAI Agent Builder supplied the visual workflow editor, ChatKit supplied the embedded chat interface, and Evals supplied the testing environment. Your migration must account for those components separately.

Think of it as a furnished workspace: the room supplied the editing and running environment, but the workflow describes the work you need done. Losing that room means finding another place to operate. It does not mean discarding every tool or switching model providers.

The retirement dates come from OpenAI's deprecations page; the continuing runtime choices are documented in its agents overview. The notices were announced on June 3, 2026. October's read-only change and November's shutdown are different events.

The Responses API is not deprecated by this announcement. A Responses application can still be affected if its requests reference reusable prompt objects. The endpoint you call and the saved resources it depends on are different things.

Who Needs to Act Now?

Act now if any live workflow depends on Agent Builder, managed Evals, or saved prompt objects. The right destination depends on where the workflow belongs, not which product name sounds most similar.

Builders shipping a customer-facing app: export the workflow and assign an owner for its runtime. A support chatbot that must run inside your product needs a working application service after the hosted workflow closes. Moving only the visual interface leaves the important dependency unresolved.

Operators running internal work: consider ChatGPT Workspace Agents when the job can live in your company's ChatGPT workspace. A recurring account brief or internal research assistant is a stronger fit than a public website integration. Have the process owner judge the output and the workspace administrator review access.

Buyers about to start: choose a continuing path now. For a vendor proposal based on Agent Builder, ask for the post-shutdown runtime, migration deliverable, and acceptance criteria before approving the implementation. A demo that works during the transition window does not answer those questions.

Existing SDK or direct Responses users: you may have no Builder migration to perform. Check your evaluation pipeline and saved prompt references before declaring the application unaffected. If those dependencies are absent too, this announcement alone is not a reason to rewrite the app.

Export Agent Builder Workflows to the Agents SDK

OpenAI Agent Builder exports workflow code, which gives you a starting point for an application you operate. Its current migration guide offers Python and TypeScript exports. Keep the complete export before changing anything, so you can compare later edits with the original. Migration instructions.

1. 1

   Copy the workflow export

   Open the workflow, select Code in the top navigation, choose Agents SDK, and select Python or TypeScript. Copy the whole export into your repository. Record which deployed workflow it replaces.

2. 2

   Make the application runnable

   Install and configure the matching SDK using its current quickstart. Connect the tools and credentials the workflow expects. Assign deployment, storage, and runtime ownership before treating the export as a replacement service.

3. 3

   Prove the business behavior

   Replay representative inputs against the replacement. Check the chosen branch, tool arguments, approval behavior, and final result. Deploy only after the replacement meets the workflow's acceptance criteria.

The OpenAI Agents SDK gives your application control over deployment, storage, tools, and approvals while it runs the agent loop. The SDK overview is the continuing implementation reference.

The SDK's wall is operational ownership. A copied file does not supply your production deployment, restart behavior, or persistent state. For a support workflow, define what happens when an order lookup times out, when an approval is denied, and when the same request arrives again. Those are acceptance cases to implement and verify, not promises made by an export button.

Use the OpenAI Agents API vs Agents SDK comparison if you are also reconsidering who should operate the runtime. OpenAI's current Builder migration guide names the SDK and Workspace Agents; do not treat the separate managed Agents API as an automatic import destination.

Move No-Code Work to ChatGPT Workspace Agents

ChatGPT Workspace Agents is the documented route for natural-language workflows shared inside a team. OpenAI's migration guide requires a Business, Enterprise, or Edu workspace with agent access and permission to create agents. Export the Builder workflow, paste that code into the workspace agent builder, and ask it to recreate the workflow. Workspace migration requirements.

OpenAI's workspace agent walkthrough shows the relevant operating model: connected apps, reusable instructions, scheduled work, and sharing inside a workspace. That suits an internal account-brief workflow whose owner can inspect the finished brief.

The hard limit is behavior fidelity. OpenAI warns that workflows requiring strongly deterministic behavior, where the same conditions must follow the same prescribed path, may not transfer faithfully. Review connections and permissions, use Preview with representative inputs, and select Create after validation. This is a recreation process, not a guaranteed conversion of every node and edge.

For example, an internal research assistant can tolerate different research routes if the final brief meets its requirements. A workflow that must always obtain approval before changing a customer record needs explicit verification of that boundary. If your business depends on exact branching, my recommendation is to retain that logic in application code and validate the SDK path first.

Keep ChatKit, Replace the Hosted Backend

ChatKit remains useful as your chat front end; your application must supply the replacement server behind it. OpenAI directs new projects and existing Builder-backed apps toward a custom server integration. Existing hosted workflows are supported only during the transition window. ChatKit transition guidance.

This changes the work behind the chat box. In the continuing architecture, a browser sends requests to your ChatKit server, the server runs your agent implementation and tools, and results return to the interface. Keeping the same visual experience does not preserve the old hosted execution path.

The advanced integration guide identifies the implementation pieces: a ChatKitServer, an HTTP endpoint, durable storage for production threads and messages, a file store if you support uploads, and server context for identity and permissions. It also provides the connection to the Agents SDK.

For a customer-support app, acceptance should include more than receiving a reply. Confirm that the signed-in customer sees only the intended conversation, attachments remain usable where required, permitted tools execute, and escalation still reaches the right destination. Plan how existing conversation history will be handled; do not assume it follows the workflow export.

The guide to building a chatbot with ChatKit covers the wider product responsibilities. For this migration, the decisive question is whether a real conversation completes through the replacement backend with the expected access and action rules.

The Dated Migration Checklist

Save the evaluation baseline first, then move execution and prompts before the shutdown. A baseline is the set of examples and acceptable outcomes you use to decide whether the replacement works. Without it, you can move code and still lose the behavior your users rely on.

Before October 31, 2026: Export and Preserve Evals

Assign an evaluation owner now. Preserve the test inputs, expected outcomes, grader definitions, prompt versions, model settings, and important existing results needed to understand the current behavior. Export what your existing workflow exposes; where no export is available, copy the configuration and recreate it outside the platform.

October 31 is the read-only milestone, not a documented final export deadline. Finishing the backup before then is the recommended project deadline here, so you have a usable baseline while migrating.

OpenAI's Evals migration guide recommends Promptfoo and describes manual recreation of prompts, providers, test cases, and assertions. It does not require an OpenAI Evals export feature. Re-run and validate recreated graders before relying on their scores; OpenAI warns that scoring may differ across systems.

A useful handoff is a runnable evaluation suite plus an explanation of what counts as failure. A screenshot of a previous score is supporting history, not a replacement for that suite.

Before November 30, 2026: Move Workflows and Prompt Objects

Finish these tasks before the scheduled shutdown:

• Deploy the exported SDK workflow or validate and create its Workspace Agent replacement.
• Move every Agent Builder-backed ChatKit app onto the replacement server.
• Move reusable prompt content out of managed prompt objects.
• Run the replacement evaluation process and investigate behavior changes.
• Remove production dependencies on retiring services and assign an owner to the new operating path.

Prompt objects deserve a separate search through your application. A workflow can run successfully while another feature still depends on a saved prompt ID. OpenAI's prompt migration guide moves prompt text into source code, replaces variables with function arguments, and passes generated messages through Responses API input instead of the managed prompt object. Version the prompt alongside the application and retain a stable shared prefix before dynamic values where practical.

For a billing assistant, the deliverable is the complete instruction set, its variable substitutions, and the expected responses to representative cases. Copying only the visible paragraph while losing a variable or choosing the wrong saved version is an incomplete migration.

Set your own cutover date earlier than the vendor deadline, with room to investigate a failed test. Keep a documented fallback during the remaining transition period; a fallback that still relies on a retired service cannot protect you after shutdown.

What's Overhyped or Easy to Misread

“AgentKit is shutting down” is too broad to guide an engineering decision. The notices name products and saved resources. The SDK, ChatKit, Responses API, and separate Agents API have continuing documentation; the affected hosted services need migration.

An export is not proof of parity. A generated replacement can produce a convincing answer while skipping the branch that required approval. Compare observable behavior with the original requirements, not whether the new diagram or code looks familiar.

ChatKit survival does not preserve Builder hosting. The interface can remain while the service doing the work changes. Budget engineering ownership accordingly, without assuming a new frontend is required.

The Evals shutdown does not remove the need for evaluations. It moves responsibility for maintaining your tests. Nor does moving a prompt into code retire the Responses API; it changes how your application supplies instructions to it.

Questions About Using AgentKit Now

Is the OpenAI agent kit free?

Do not base a migration on the assumption that AgentKit is a single free product. Choose a continuing runtime first, then check its current model, tool, and infrastructure charges. The shutdown notices do not establish a universal migration price.

How to use OpenAI agent?

For an existing Agent Builder workflow, export to the Agents SDK or recreate it as a Workspace Agent. For a new app, choose a current runtime and test the tools and permissions your job requires. A ChatKit interface needs a server-side implementation.

What can you do with OpenAI agents?

Agents can use tools and work through multi-step tasks. A useful migration starts with the specific job you already depend on, such as preparing an internal brief or answering an authenticated support request, and preserves its required behavior on the new runtime.

Your Monday Move

Name a migration owner and choose a representative live workflow. Save its export, prompt content, and evaluation baseline; then run a replacement end to end. Use the result to estimate the remaining inventory. Your first deliverable should be evidence that a required job completes with the right permissions and outcomes, followed by the dated plan for moving the rest.

For practical updates on building and operating AI workflows, join the newsletter.