OpenAI Agent Builder and AgentKit: what shuts down Nov 30

OpenAI Agent Builder, the visual canvas of AgentKit, shuts down on November 30, 2026. OpenAI announced it on June 3. Both ways out start with one step: export each workflow to Agents SDK code. Then run that code yourself, or turn it into a Workspace Agent in ChatGPT. ChatKit, the Agents SDK and the Responses API stay. Evals goes too, read-only from October 31.

That's the short version. The long one matters if a customer-facing bot runs on an Agent Builder workflow, because the export button lives inside Agent Builder. After November 30 there's nothing left to export.

AgentKit a year later: what's left

OpenAI launched AgentKit on October 6, 2025, as one box of parts for building agents: a visual canvas, an embeddable chat window, evals, a connector hub and guardrails.

A year on, OpenAI's docs no longer describe AgentKit as a product suite. The parts, and the tools around them, live on separately. Two of them close on the same day.

Part	What it does	Status	Where to move
Agent Builder	Visual canvas for multi-agent workflows	Shuts down November 30, 2026	Agents SDK (code) or Workspace Agents in ChatGPT
ChatKit	Embeddable chat interface	Stays. Workflows hosted in Agent Builder are in a transition window	ChatKit on your own server
Evals: datasets, trace grading, prompt tuning	Tests agent output	Read-only October 31, shuts down November 30	Promptfoo, per OpenAI's deprecation note
Reusable prompts (v1/prompts)	Stored prompts called by ID	Shuts down November 30	Prompt text in your own code
Agents SDK, Python and TypeScript	Code-first agent loop with tools and handoffs	Stays, updated this year with sandboxes and memory control	None
Guardrails	Open-source safety layer	Open-source libraries, still in preview	None
Connector Registry	Admin hub for data connections	Not in current OpenAI docs	Confirm with OpenAI before you rely on it
Assistants API	Older API for agents with threads	Shut down August 26, 2026	Responses API and Conversations API

What stops working on November 30

Agent Builder itself. No more editing or publishing workflows.

Then everything that leans on it. A ChatKit window that calls a hosted workflow, the kind with an ID that starts with wf_, runs on Agent Builder. OpenAI's ChatKit docs say these workflows are in a transition window and send all new work to a server you run. OpenAI's docs say nothing about these embeds after November 30. Treat them as dead on December 1.

Starting something new? Skip Agent Builder and build on your own server from day one.

Prompt objects close the same day. If your code calls a stored prompt by ID, that call fails.

Evals go in two steps. On October 31 your existing evals freeze. On November 30 the dashboard and the API close. If your release process runs a regression set there, copy the test cases out now.

How to export an Agent Builder workflow to code

OpenAI built an export into Agent Builder:

Open the workflow in Agent Builder.
Click Code in the top navigation.
Choose Agents SDK, then TypeScript or Python.
Copy the full export into your repository and commit it.

Do it for every workflow, including the ones you think are dead. A copy costs nothing. A missing one costs a rebuild from memory.

Now the fine print. OpenAI's migration guide says the export "does not convert your workflow graph" and doesn't promise every behavior carries over. So run the exported agent next to the old one before November 30, while you can still compare answers. Check these first:

Routing. Check that every if-else branch from the canvas made it into the code. Read it line by line.
Escalation. Find where the workflow hands a conversation to a person and test it with a real ticket.

Then the plumbing. Re-add guardrails if they didn't come through, paste any prompt pulled by ID into the code, confirm file search points at the same vector store IDs, and re-check API keys and what each tool may touch.

Then host it. The SDK runs on your server, and the ChatKit Python SDK connects your chat window to it. OpenAI keeps a customer service demo on the Agents SDK, openai-cs-agents-demo, with a ChatKit front end. Read it before you design your own.

Agents SDK or Workspace Agents

OpenAI's migration guide names two destinations, and neither is a drag-and-drop canvas. Pick by who keeps the agent running.

The Agents SDK fits when engineers own the agent. Routing lives in code, and so do the tests. You pay for model tokens and tool calls.

Workspace Agents fit internal helpers. You build them in plain language inside ChatGPT. You need a Business, Enterprise or Edu workspace with permission to create agents. OpenAI warns that workflows "with strong determinism at their core may not migrate faithfully" there. A customer-facing support bot is exactly that kind of workflow: refund rules and a fixed escalation path. For anything that talks to customers, take the SDK.

One name trap. OpenAI also runs an Agents API (v1/agents), in public beta since September 10, 2026. It runs agents on a Codex setup that OpenAI manages. It's a separate product from the Agents SDK, and the migration guide doesn't list it as an Agent Builder replacement.

Is OpenAI Agent Builder free? What the move costs

OpenAI publishes no separate price for Agent Builder. At launch it said the AgentKit tools came with standard API model pricing, so you paid for tokens and tool calls. That stays the same on the SDK. Workspace Agents come with the paid ChatGPT plan instead.

The tool calls you keep paying for:

Web search: $10 per 1,000 calls, plus search content tokens at model rates.
File search: $2.50 per 1,000 calls, Responses API only.
File storage: 10 cents per GB per day after the first free GB.
ChatKit upload storage: 10 cents per GB per day after 1 GB free each month.

A quick check. A support bot handles 5,000 conversations a month and runs file search twice per conversation. That's 10,000 calls, or $25 a month. Tokens come on top and depend on the model.

The line that grows is engineer time. Agent Builder hosted the workflow for you. After the move, someone on your team owns the server, the deploys, the logs and the eval set that used to sit in OpenAI's dashboard.

The support-team question: rebuild, or move the queue

Agent Builder made a support bot look like an afternoon of dragging nodes. Then OpenAI retired the Assistants API on August 26 and set Agent Builder to close on November 30. That's 96 days apart, and both rebuilds land on whoever owns the code.

Before you rebuild, answer four questions:

Who keeps the bot's knowledge in sync with the help center?
Who reads the conversations the bot got wrong, every week?
How does the bot hand a ticket to a person inside your helpdesk, with the context attached?
Who re-tests everything when OpenAI changes the model underneath?

If an engineer owns the bot and has the hours, export it and rebuild on the SDK. If support owns it, write a weekly hour count next to each answer. That number is the real cost of running it yourself, and any vendor invoice has to beat it.

Here's ours to compare against. CoSupport AI Agent answers inside Zendesk, Intercom or Freshdesk and passes what it can't resolve to your agents in the same queue. It costs $0.50 per resolved AI conversation (no customer reply within 72 hours after the AI answer), and the first 1,000 AI responses are free.

FAQ

Is OpenAI Agent Builder being discontinued?

Yes. OpenAI announced it on June 3, 2026, and Agent Builder shuts down on November 30, 2026, with the Agents SDK and Workspace Agents in ChatGPT as the two migration paths.

Is OpenAI AgentKit still available?

AgentKit was the October 2025 launch name for a set of tools, and OpenAI's docs no longer describe it as a product suite. ChatKit stays, and Guardrails lives on as open-source preview libraries. Agent Builder and the Evals platform shut down on November 30, 2026.

Will ChatKit stop working?

ChatKit stays. A ChatKit window that runs a workflow hosted in Agent Builder depends on Agent Builder, and OpenAI calls these workflows a transition case. Move that workflow to your own server with the Agents SDK before November 30. ChatKit connected to your own backend keeps working.

How do I export an Agent Builder workflow?

Open the workflow, click Code, choose Agents SDK, pick TypeScript or Python and copy the export into your repository. The export doesn't convert the workflow graph one to one, so test the new agent against the old one before November 30.

Is OpenAI Agent Builder free?

OpenAI publishes no separate price for Agent Builder: you pay model tokens plus tool calls, such as file search at $2.50 and web search at $10 per 1,000 calls. The same prices apply on the Agents SDK.

What replaces OpenAI Evals?

OpenAI's deprecation notice points to Promptfoo. Existing evals become read-only on October 31, 2026, and the dashboard and API shut down on November 30, so move your test cases before then.