twimco.Let’s talk ↗︎
Menu
Field notes / AI in practice

OpenAI's Agent Builder shuts down November 30. Record what your workflows do today, before you move them

After November 30 there's no old version left to compare against. That's the real deadline: two weeks to write down what each Agent Builder workflow actually does, while it still runs.

Editorial engraving of three colleagues in a sunny loft comparing two side-by-side monitors, one pointing with a pencil while another types and a third holds a paper calendar.

OpenAI's docs say Agent Builder is scheduled to shut down on November 30, 2026. You can export each workflow as Agents SDK code or rebuild it as a ChatGPT workspace agent, but OpenAI says the export doesn't convert the workflow graph or guarantee identical behavior. Before moving anything, run a set of real inputs through the current workflow and save the outputs, so the new version has something to be checked against after the old one is gone.

The deadline behind the deadline

OpenAI's documentation says Agent Builder, the visual tool for laying out multi-step agent workflows, is scheduled to shut down on November 30, 2026. Existing users can keep using it during the transition window. ChatKit, the kit many teams used to put those workflows in front of customers, stays available. Forkast reports the shutdown was first announced on June 3, so for some teams this is old news that slid down the list.

The date that matters more is quieter. Once Agent Builder is gone, you can't run the old version of a workflow again. You can't ask it what it would have done with last Tuesday's awkward customer message. So the real job for the next two weeks isn't only moving workflows. It's writing down what each one actually does while it still answers.

What OpenAI says the move involves

The migration guide is short and candid. Open the workflow, choose Code, choose Agents SDK, pick TypeScript or Python, and copy the whole export. From there you have two paths.

The first is to keep going with the Agents SDK: paste the export into an application you run, install the matching SDK, configure it, and test it in your own runtime. The guide says to validate configuration, tools, authentication, permissions and deployment before going live.

The second is to turn the export into a ChatGPT workspace agent. That needs a Business, Enterprise or Edu workspace with access to workspace agents and permission to create them. You create an agent, paste the exported code with a short prompt asking ChatGPT to convert it, and review anything it flags as needing changes. Then you configure apps, tools, skills and connection permissions, preview it with representative inputs from the original workflow, compare the results, and only then press Create.

One line in the guide is worth printing out: the export "does not convert your workflow graph or guarantee that every behavior transfers unchanged." It also warns that workflows with "strong determinism at their core" may not move faithfully to a workspace agent. If your workflow is a strict sequence, check this step, then that one, never skip, the SDK path is the safer home.

Sequence of work

Two weeks, worked backward from November 30

  1. Mon Nov 16 to Wed Nov 18

    List every published workflow, who uses it and where its ID is called from.

  2. Thu Nov 19 to Sat Nov 21

    Pick real inputs for each one and save the current outputs.

  3. Mon Nov 23 to Wed Nov 25

    Export to Agents SDK code; rebuild or convert on the chosen path.

  4. Thu Nov 26 to Fri Nov 27

    Run the saved inputs through the new version; mark every difference.

  5. By Mon Nov 30

    Switch callers over, keep the bench file, retire the old ID.

Illustrative schedule built from Monday, November 16; the shutdown date is from OpenAI's Agent Builder documentation.
The old workflow is the only reliable description of what it does, and it stops answering on November 30.

Build the bench before you move anything

Here's the habit we'd put in place first, and it takes an afternoon per workflow. Open a spreadsheet with four columns: input, what the current workflow did, what the new version did, and whether the difference matters.

For the first column, use real inputs, not invented ones. Agent Builder keeps traces of past runs and lets you grade them, so pull ten from there: a few ordinary ones, the long rambling one, the one in another language, the one with a missing order number, the one that should have been refused. Run each through the current workflow and paste the full output into the second column. Note which tools it called and in what order, not just the final text.

Then, and only then, export and rebuild. Run the same inputs through the new version and fill in the third column. Most rows will match closely enough. The ones that don't are your real migration work: a branch that no longer fires, a tool that now runs before a check instead of after, a refusal that turned into an answer. The old workflow is the only reliable description of what it does, and it stops answering on November 30.

Keep the file afterwards. It becomes the regression test for the next model change, and there will be one.

The question nobody has answered yet

Forkast points out something the documentation doesn't settle: whether workflows published in Agent Builder and called by ID through ChatKit will keep running after the shutdown. ChatKit staying available doesn't by itself mean your old workflow ID will. Don't plan on it. Search your code and your site for every place a workflow ID is passed in, so you know exactly which screens a customer would see break, and switch those to the migrated version before the 30th.

While you're searching, write down who owns each workflow. Visual builders tend to collect workflows made by someone who has since moved on. If nobody can say what a workflow is for, that's a fair candidate for retiring rather than migrating.

If you only have an afternoon

Not every team has two spare weeks before Thanksgiving. If that's you, shrink the job rather than skipping it.

Start with the one workflow customers touch. Pick five real inputs instead of ten: two ordinary, one awkward, one that should be refused, one with something missing. Save the current outputs today, even if you don't export anything until next week, because that's the part you can't redo later. Then export the code and store it with the bench file in the same folder, dated, so whoever picks it up knows exactly what "working" looked like on November 16.

Next, put a note in the calendar for the morning of December 1 to open the screens that used each workflow and check they still respond. If something breaks, you'll know before a customer tells you, and you'll have the saved outputs to rebuild against.

Decision guide

Which path for which workflow

ConsiderationAgents SDK in your own appChatGPT workspace agent
Best whenThe workflow runs inside your product or a back-office systemPeople in your workspace start the job and read the result
NeedsA developer, a runtime, keys and deploymentA Business, Enterprise or Edu workspace and permission to create agents
Strict, step-by-step logicCode keeps the order you writeOpenAI warns strongly deterministic workflows may not migrate faithfully
What you checkRuntime config, tools, auth, permissions, deploymentApps, tools, skills, auth and connection permissions
Decision framework drawn from OpenAI's migration guide (source: developers.openai.com, fetched Nov 16, 2026); examples are illustrative.

Choosing the path, workflow by workflow

You don't have to pick one path for everything. A customer-facing assistant embedded in your site belongs in code you control, with the SDK. An internal helper that a few colleagues start by hand and read in ChatGPT may be happier as a workspace agent. Make the call per workflow, using the bench results: if the workspace version keeps failing the same rows because it improvises where the old one followed steps, that's your answer.

Either way, keep the same safety habits you had before. OpenAI's guide says so directly, especially where an agent can reach private data or act through connected tools. A migration is a good moment to remove a permission nobody needed, and to write down, in one line per workflow, who gets told when it fails.

If you'd like help moving agent workflows into code you own, with tests that outlive the move, that's what we do in AI applications and agents. Talk to us.

Related work: Spin:Market

twimco.

We build alongside the people doing the work.
Bring us the problem you’re thinking about.

Related work

Spin:Market ↗︎