OpenAI Agent Builder shuts down on 30 November 2026: export your workflows, keep ChatKit, move prompt objects

Agent Builder: scheduled to shut down on November 30, 2026
v1/prompts: scheduled to shut down on November 30, 2026
# OpenAI's notice, not an error message. OpenAI hasn't published what a workflow ID (wf_...) or a prompt ID (pmpt_...) returns after that date. We add the real text here on 30 November.

Short answer: OpenAI shuts down Agent Builder and prompt objects (v1/prompts) on 30 November 2026. Export each workflow now as Agents SDK code. Move prompt text into your own code, and run ChatKit against your own backend. ChatKit itself stays, but OpenAI tells you to leave the hosted Agent Builder path before the shutdown.

Checked on 6 November 2026 against OpenAI's deprecations page and its Agent Builder, prompt object and ChatKit guides.

Why it happens

OpenAI told developers on 3 June 2026. Its deprecations page lists three shutdowns for 30 November 2026: Agent Builder, the Evals dashboard and API (existing evals went read-only on 31 October), and the v1/prompts API with reusable prompt objects. ChatKit stays available, but the path that embeds a workflow by its wf_ ID depends on Agent Builder: OpenAI's ChatKit guide says to use that path only if you already have such a workflow, and to use the advanced integration, on your own server, when migrating before the shutdown. How we verified it (a dated vendor change): every step comes from OpenAI's documentation, read on 6 November 2026. We have no OpenAI API key here, so we ran no export and no live call. What we could test that day: openai-chatkit 1.6.5 and openai-agents 0.24.0 are on PyPI and @openai/agents 0.21.0 is on npm, the openai-chatkit wheel contains the ChatKitServer class with its respond method, and POST /v1/chatkit/sessions without a key answers 401.

Fix

List what you have. In Agent Builder, note every workflow you still use. In your code and settings, search for workflow IDs and prompt IDs:
```
git grep -nE "wf_[0-9a-f]+|pmpt_|prompt_id|chatkit/sessions"
```
Export each workflow now, while the dashboard works: open it in Agent Builder, select Code, then Agents SDK, choose TypeScript or Python, and copy the complete export into your repository. OpenAI says this doesn't convert your workflow graph and doesn't guarantee that every behavior transfers unchanged, so write down what each workflow should do.
Choose where it runs. On your own server: pip install openai-agents or npm install @openai/agents zod, set OPENAI_API_KEY, and check configuration, tools, authentication and permissions yourself. As a ChatGPT workspace agent: this needs a ChatGPT Business, Enterprise or Edu workspace. Paste the export into a chat, ask it to convert the workflow, review what it flags and test in Preview. OpenAI warns that workflows with strong determinism may not migrate faithfully.

Move prompt objects into code. OpenAI's guide names no export tool, so copy the text of each prompt version you use from the dashboard. Turn each variable into a function argument and send the messages as input. The new call must name the model itself:

# before: a saved prompt object, the form OpenAI shuts down
response = client.responses.create(
    prompt={"id": "pmpt_123", "version": "1",
            "variables": {"customer_name": "Acme", "issue": "billing question"}}
)

# after: the same text in your own code, one function per prompt
def support_prompt(customer_name, issue):
    return [
        {"role": "system", "content": "TEXT COPIED FROM THE PROMPT OBJECT"},
        {"role": "user", "content": f"Customer name: {customer_name}. Issue: {issue}."},
    ]

response = client.responses.create(model="YOUR_MODEL", input=support_prompt("Acme", "billing question"))

Keep static text first and dynamic values last, so prompt caching keeps working, and use git for versions.

Using ChatKit? If it starts sessions with a wf_ ID, move it to the advanced integration: pip install openai-chatkit, subclass ChatKitServer, override respond, expose it as an endpoint (OpenAI's example uses FastAPI), add a Store for threads and messages and a file store if you allow uploads, then point your front end at the new endpoint.
Test with real inputs before 30 November and compare with the old workflow. Then repoint your app and remove every wf_ and pmpt_ reference.
Already past 30 November with no export? OpenAI's guide shows no other way to get one, so rebuild the workflow with the Agents SDK from your notes.

Built it in Agent Builder and have no developer to move it? Send the exported code or your list of prompts. I turn one workflow, or all prompt objects in one repository, into code you own: a pull request with deploy notes for a low-cost host, within 48 hours. I check that it installs, starts and runs its tools with the model calls stubbed, because I don't ask for your API key. You run the last check with your key and deploy it.

Get it moved: $49

Pay by card, no account needed. Still failing after you merge and run it with your key: full refund within 30 days.

How it works: 1. Pay on Gumroad; the checkout asks for the repository. 2. Invite my GitHub user, Donrulezz, to that repository only (private is fine), or email a zip with the failing log to hello@deadlinefix.com. 3. You get a pull request with a list of every change; you review, merge and deploy. I never ask for cloud passwords or production keys.

Who does the work: me, Ashish, the owner of this site, based in India. My public code and tools are on GitHub: github.com/Donrulezz.

Watch out

Not confirmed by OpenAI: what a hosted ChatKit session returns after 30 November. OpenAI calls hosted workflows a "transition window" and says to move before the shutdown. A forum user expects that backend to go away. Plan as if it stops.
Also not confirmed: whether the export still works after the date. OpenAI shows it only as a step inside Agent Builder, and a forum user says it works only while the dashboard runs. Do it now.
The export needs somewhere to run, and that costs money. Forum users reported in June 2026: an always-on VM at about $15 to $30 a month, cold starts of just under 20 seconds on one cheap host, and about 3 seconds on a 1024 MB AWS Lambda. These are their numbers, not our tests.
One forum user found that openai-guardrails 0.2.1 doesn't install cleanly on Python 3.14. If your export imports it, check your Python version first.
Evals goes the same day. Existing evals have been read-only since 31 October, and OpenAI points to Promptfoo.
An exported workflow that reads private data or acts through tools needs the same safety review as the original, as OpenAI's guide says.

Sources

Did this fix it?

Other errors breaking soon

All "breaking soon" fixes