How to Migrate OpenAI Agent Builder Workflows Before the November 30 Shutdown

Introduction

OpenAI Agent Builder arrived on October 6, 2025, as the visual centerpiece of AgentKit at DevDay: a drag-and-drop canvas for building multi-step agent workflows without writing code. Less than eight months later, on June 3, 2026, OpenAI told developers using it that the product was being deprecated. The shutdown date is November 30, 2026, which leaves 23 days from today. This tutorial walks you through the rescue: inventory every workflow, export it as Agents SDK code, run it in your own app or rebuild it as a ChatGPT Workspace Agent, move your evals to Promptfoo, and inline your reusable prompts before the canvas goes dark.

Quick answer

Export each workflow before November 30, 2026: open it in Agent Builder, select Code, choose Agents SDK, pick TypeScript or Python, and copy the code. Then run it with the Agents SDK, or rebuild it as a ChatGPT Workspace Agent.

Evidence

OpenAI’s own deprecations page carries the shutdown in plain language: “On June 3, 2026, we notified developers using Agent Builder that the product is being deprecated. ChatKit remains available.” The same page lists three shutdowns for November 30, 2026: Agent Builder, the Evals dashboard and API, and the v1/prompts API for reusable prompt objects. The Evals platform already went read-only on October 31, 2026, so that deadline has passed while the other two are still open.

The migration destination is not a niche project. The Python Agents SDK repository, openai-agents-python, had 30,536 stars on GitHub as of November 7, 2026.

The most important sentence in OpenAI’s migration guide is also the most honest one: “This process does not convert your workflow graph or guarantee that every behavior transfers unchanged.” Treat the export as a faithful code draft of your agents, instructions, and tools, not as a one-to-one transplant of the workflow you built on the canvas. Plan your migration around that sentence, and nothing in this tutorial will surprise you.

Prerequisites

• Access to every Agent Builder workflow you own, in the OpenAI Platform, before November 30, 2026.
• An OpenAI API key, if you choose the Agents SDK route.
• Python 3.10 or newer, or Node.js 18 or newer, for the Agents SDK route.
• A ChatGPT Business, Enterprise, or Edu workspace with workspace agents enabled, if you choose the Workspace Agent route.
• A list of every connected app, tool, and credential each workflow touches, because the export does not carry those over.

Step 1: Inventory everything before you touch the export button

Open Agent Builder and list every workflow, including drafts and ones you have not opened in months. For each one, write down four things: what triggers it, which tools and connected apps it calls, which model and instructions each agent uses, and who is allowed to run it. Screenshot the canvas, because after November 30 you cannot open it to check a detail you forgot.

Do the same for the two companion products. List every evaluation in the Evals dashboard, noting its dataset, prompts, and graders; evals have been read-only since October 31, so you can still read them but cannot edit them. Then list every reusable prompt object, because the v1/prompts API shuts down on November 30 too, and any application that fetches a prompt by ID will break that day.

Rank the list by business impact, not by size. A three-node workflow that files support tickets every day outranks a twenty-node demo nobody runs.

Step 2: Export each workflow as Agents SDK code

OpenAI documents exactly one export path, and it takes four clicks. Open the workflow in Agent Builder, select Code in the top navigation, select Agents SDK in the code dialog, then select TypeScript or Python and copy the complete export. Save each export as a file in your repository immediately, named after the workflow, for example support-triage-agent.py.

Pick the language your team already runs. If the workflow will live inside an existing Python service, export Python; if it will live in a TypeScript backend, export TypeScript. Export every workflow on your inventory before you start testing any of them, so one broken workflow cannot block the rest of the rescue.

Step 3: Run the export with the Agents SDK

The Agents SDK route is the right choice when the workflow runs inside your own product or API, or when it needs strict, repeatable steps. Install the matching SDK first. For Python, run pip install openai-agents. For TypeScript, run npm install @openai/agents zod. Both commands come from the SDK repositories’ own quickstart instructions.

Paste the exported code into your application and set your API key as the OPENAI_API_KEY environment variable. The Agents SDK overview in OpenAI’s documentation describes the framework as a small set of primitives: Agents, handoffs between agents, and Guardrails that validate inputs and outputs. Your export already expresses your workflow in those terms.

Run the workflow with a real input from your production logs, not a toy prompt. A minimal smoke test in Python looks like this:

from agents import Agent, Runner

agent = Agent(name="Assistant", instructions="You are a helpful assistant")
result = Runner.run_sync(agent, "Write a haiku about recursion in programming")
print(result.final_output)

That snippet is the SDK’s own quickstart shape, and it works as a harness check: if it runs, your key, install, and runtime are healthy. Then run your exported agent the same way and compare its output with the original workflow on the same input, side by side.

Before you deploy, validate the four things the migration guide names: runtime configuration, tools, authentication, and permissions. The export does not carry your connected apps or credentials, so re-attach each tool, re-enter each secret through your own secret store, and re-check who can invoke the agent. If your workflow had approval steps or guardrails on the canvas, re-create them as SDK guardrails and test that they block a bad input. For a deeper grounding in the SDK, our OpenAI Agents API tutorial walks through a first working agent end to end.

Step 4: Rebuild as a ChatGPT Workspace Agent instead

Choose this route if your team builds agents in natural language and shares them inside ChatGPT, and you have a Business, Enterprise, or Edu workspace with workspace agents enabled and permission to create agents.

In ChatGPT, create a workspace agent, then paste your exported code into the chat with this prompt:

Please help me convert this workflow into an agent:
<paste your exported code here>

Review any behavior the builder flags as needing changes before you continue. Then, before you select Create, work through OpenAI’s pre-flight list: review the generated instructions and capabilities, configure required apps, tools, skills, authentication, and permissions, select Preview, test representative inputs from the original workflow, and compare the result with the original’s expected behavior. Select Create only after you have validated the migrated agent.

Two warnings from the migration guide apply here specifically. Workflows with strong determinism at their core may not migrate faithfully to a workspace agent, because a conversational agent does not execute a fixed graph the way the canvas did. And connected apps, authentication, publishing, and permissions require separate review in ChatGPT; none of that rides along with the pasted code. If your workflow is a strict multi-step pipeline with approvals, the Agents SDK route in Step 3 is usually the safer home.

Step 5: Move your evals to Promptfoo

OpenAI does not offer an Evals export. The supported path, documented in OpenAI’s cookbook guide on moving from OpenAI Evals to Promptfoo, is to recreate each evaluation manually in Promptfoo, the open-source evaluation CLI. Promptfoo centers on a portable configuration file instead of a dashboard, so your evals end up living alongside your application code, which is an improvement you keep after the migration.

Install Promptfoo, then recreate each evaluation’s prompts, provider, test cases, and assertions in a promptfooconfig.yaml file. Map each OpenAI Evals grader to a Promptfoo assertion, and each dataset to test cases in the config. Run the recreated evaluation and open the results viewer:

promptfoo eval -c promptfooconfig.yaml --no-cache
promptfoo view

If the configuration has an error, validate it before retrying with promptfoo validate config -c promptfooconfig.yaml. After your first run, compare the recreated prompts, assertions, and outputs with the original eval before you rely on the new one for regression decisions. Similarity-based scoring may not produce identical numerical results across the two systems, and any manually recreated grader, especially an LLM-as-a-judge grader, should be validated before you trust it.

Step 6: Inline your reusable prompts into code

The v1/prompts API and reusable prompt objects shut down on November 30, 2026, on the same day as Agent Builder. OpenAI’s deprecations page is direct about the fix: move reusable prompt content into your application code.

For each prompt object, copy its text into a constant or a configuration file in the application that uses it, and replace the API call that fetched it by ID with a reference to that local copy. Search your codebase for prompts/ API calls and prompt IDs, because a prompt fetched at runtime from a dead endpoint fails on December 1 even if the rest of your migration is perfect.

Step 7: Test side by side, then cut over by November 25

Run the old workflow and the new implementation on the same five to ten inputs from your real traffic, and compare outputs, tool calls, and latency. Small differences in wording are normal, because the export is a code draft and the runtime is different. Differences in which tools ran, which data was touched, or which approvals fired are not normal: each one is a bug to fix before cutover.

Set your own deadline five days early: finish the cutover by November 25. The shutdown is November 30, and the last week is the wrong time to discover that a credential expired or a workspace permission was missing. After cutover, confirm the old workflow is idle, then revoke any credentials that only the canvas used. If a workflow turns out to be unmigratable in the time left, document what it did and who depended on it, and rebuild it as a fresh Agents SDK project after the shutdown rather than leaving a silent gap.

For teams that already orchestrate agents outside OpenAI, the same export-first discipline applies to the rest of the stack: our complete n8n guide covers building durable agent workflows in n8n, a practical home for the scheduling and integration glue that the canvas used to provide.

Troubleshooting

The export fails with “Unknown tool type: web_search”

If any agent in the workflow has the built-in Web Search tool, Agent Builder refuses the export. Remove the Web Search tool from the workflow, export again, and add web search back in the Agents SDK with the WebSearchTool, or route search through your own tool in the new code.

The exported code runs, but behaves differently from the canvas

This is expected, and OpenAI says so in the guide: the export does not convert your workflow graph. Rebuild the control flow by hand in the SDK code: check the order agents run, the conditions that route between them, and any loops or retries the canvas implied. Add a guardrail or an explicit handoff wherever the canvas enforced a rule silently, and re-run your side-by-side test until the tool calls match.

The workspace agent preview gives different answers than the original workflow

Conversational agents are less deterministic than a fixed canvas graph. Shorten the pasted instructions to the essential rules, move long reference material into the agent’s configured tools or skills, and test with the same representative inputs you used on the original. If the workflow is a strict pipeline, use the Agents SDK route instead.

An eval recreated in Promptfoo scores differently than the original

Similarity-based scoring does not produce identical numbers across systems, and a recreated LLM-as-a-judge grader is a new grader, not a copy. Compare the prompts and assertions line by line, run the same test cases through both, and validate the new grader on known-good and known-bad outputs before you use it for regression decisions. Do not tune the threshold to match the old number; tune it to match the behavior you actually want.

The migrated agent fails on authentication or permissions

The export carries no credentials and no permission settings. Re-enter each API key and token through your secret store, re-grant each connected app, and check the workspace permission that lets the agent act on private data. Test with the same identity the original workflow ran as, because a test run under your personal account can succeed while the production identity is still blocked.

FAQ

What exactly shuts down on November 30, 2026?

Three things: Agent Builder, the Evals dashboard and API, and the v1/prompts API for reusable prompt objects. ChatKit, the Agents SDK, the Responses API, and MCP tooling all remain available.

Does the export move my whole workflow, triggers and all?

No. The export does not convert your workflow graph and does not guarantee that every behavior transfers unchanged. You get Agents SDK code for the agents, instructions, and tools. Triggers, control flow, connected apps, authentication, and permissions need manual recreation.

Which migration route should I choose?

Choose the Agents SDK if the workflow runs inside your own product or API, or needs strict, repeatable steps. Choose a ChatGPT Workspace Agent if your team builds agents in natural language and shares them inside ChatGPT, and you have a Business, Enterprise, or Edu workspace with workspace agents enabled.

What happens to my evals after October 31, 2026?

Existing evals became read-only on October 31, and the Evals dashboard and API shut down on November 30. OpenAI recommends recreating evals in Promptfoo. There is no export button, so rebuild each eval as a promptfooconfig.yaml file and re-run it there.

Can I still create new Agent Builder workflows before the shutdown?

Existing users can keep using Agent Builder during the transition window, but starting a new build on a product with a published shutdown date is a bad bet. Anything you build now still has to be exported before November 30.

What if my workflow uses the built-in Web Search tool?

The export fails with an “Unknown tool type: web_search” error. Remove the Web Search tool from the workflow, export again, and add web search back in the Agents SDK with the WebSearchTool, or handle search through your own tool.

Conclusion

Agent Builder shuts down on November 30, 2026. The rescue is a week of focused work, not a rewrite: export every workflow as Agents SDK code, run it in your own app or rebuild it as a ChatGPT Workspace Agent, recreate your evals in Promptfoo, and inline your reusable prompts before the v1/prompts API goes dark. Finish by November 25, test against the old workflow on real inputs, then revoke the canvas-only credentials. If you are new to the SDK, start with our OpenAI Agents API tutorial.

Frequently asked questions