On this page 6 sections
The OpenAI Agent Builder deprecation takes hosted workflows offline on November 30, 2026. Export each workflow as Agents SDK code in TypeScript or Python, move its prompts into your repo, run ChatKit on your own backend and rebuild your evals in Promptfoo. Tools only your team uses can move to ChatGPT Workspace Agents instead.
OpenAI announced the deprecation on June 3, 2026, and bundled two more shutdowns with it. If you built a support bot or an internal agent on the canvas, three things you rely on stop on the same day: the workflow runtime, the hosted evals and the stored prompts. The code export runs inside Agent Builder, so it goes away with the product. The first job is to get everything out.
What the OpenAI Agent Builder deprecation removes
OpenAI’s deprecations page lists three shutdowns on November 30, all announced on June 3, 2026:
| Product | What happens | OpenAI’s replacement |
|---|---|---|
| Agent Builder | Shuts down Nov 30, 2026 | Agents SDK or ChatGPT Workspace Agents |
| Evals dashboard and API | Read-only since Oct 31, shuts down Nov 30 | Promptfoo |
v1/prompts and reusable prompt objects | Shuts down Nov 30 | Prompt text in your application code |
| ChatKit | Stays | Needs your own backend (see below) |
Three things are not on the list: the Responses API, the Agents SDK and vector stores. That matters, because it means the model calls, the agent framework and your uploaded knowledge base all survive. What disappears is the hosted layer that tied them together.
One more date sits close behind. On October 14, 2026, OpenAI announced that gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini leave the API on January 14, 2027, with gpt-6-luna as the named replacement. If any agent node in your workflow used one of those models, the exported code carries that model name with it. Change it during this migration so you don’t test everything twice in eight weeks.
Exported code needs a rebuild and retest
The export lives in the workflow itself: open it, select Code, choose Agents SDK, then pick TypeScript or Python. OpenAI’s migration guide is direct about what you get. The export “does not convert your workflow graph or guarantee that every behavior transfers unchanged.”
Treat the export as a first draft. Agent Builder’s canvas had typed nodes for agents, file search, guardrails, MCP calls, if/else branches, while loops, human approval, transforms and global state, according to the node reference. Agent nodes map cleanly onto SDK agents. The rest becomes ordinary code: an if/else node is an if statement, a set-state node is a variable or a database row, and a human approval node is a pause in your own app that waits for someone to click.
Before you rebuild anything, save four things from each workflow while the dashboard still loads:
- The exported code in both languages if you’re unsure which one you’ll use.
- A screenshot of the canvas. The graph is the only record of the branch order once the export flattens it.
- Every credential and connector the workflow uses: MCP server URLs, API keys, OAuth apps. None of that is in the export.
- The vector store IDs on each file search node.
The last one saves the most work. The file search node reads from vector stores you created on the OpenAI platform, by ID. The Agents SDK’s FileSearchTool retrieves from the same OpenAI vector stores and takes a vector_store_ids list. Vector stores are not on the deprecations list, so your help center can stay where it is. Pass the same IDs to the tool:
from agents import Agent, FileSearchTool, Runner
SUPPORT_INSTRUCTIONS = """You answer billing questions for Acme.
Only use the help center. If the answer isn't there, say so and offer a human."""
support_agent = Agent(
name="Support",
instructions=SUPPORT_INSTRUCTIONS,
tools=[FileSearchTool(vector_store_ids=["vs_your_store_id"], max_num_results=5)],
)
result = Runner.run_sync(support_agent, "How do I change my plan?")
print(result.final_output)
If your workflow is a fixed path with one or two model calls inside, rebuilding it in code is a chance to make that explicit. Our post on AI agent vs workflow design covers how to keep the outer flow in code and give the model only the step that needs it.
ChatKit stays but needs your backend
ChatKit, the embeddable chat interface, is not being shut down. The catch is in how most teams wired it. The hosted setup points ChatKit at a workflow ID from Agent Builder, and OpenAI’s ChatKit guide now says to use that path “only if you already have an Agent Builder workflow that backs your ChatKit implementation.” When the workflow goes, so does the backend behind your widget.
The replacement is what OpenAI calls the advanced integration, and it is the path the ChatKit docs recommend for anyone migrating before the shutdown:
- Install the Python server package with
pip install openai-chatkit. - Subclass
ChatKitServerand overriderespond. Inside it, run your Agents SDK agent and stream its events back. - Implement the
Storeinterface so threads, messages and files live in your own database. AFileStoreis only needed if users upload files. - Pass a context object through
server.process(body, context)to carry the user’s identity and enforce permissions. - Point the frontend’s
apiURLoption at your endpoint.
Your users see the same widget. What changes is that you now run a server, own the conversation history and handle auth. For a support bot, that’s usually a small FastAPI service and a few Postgres tables, but it is a deployment you didn’t have before, so budget for hosting and monitoring.
Prompts and evals move into your repo
The two quieter shutdowns hurt teams that didn’t know they depended on them.
Reusable prompts. If your code calls the Responses API with a prompt ID, those calls depend on v1/prompts, which shuts down on November 30. OpenAI’s guide to migrating from prompt objects says to move the prompt text into your application code, replace prompt variables with function arguments, and keep static content first and dynamic content later so prompt caching keeps working. Copy every prompt’s latest version out of the dashboard now. After that, versioning is a git commit and a pull request.
Evals. Existing evals became read-only on October 31, and the dashboard and API shut down on November 30. OpenAI recommends Promptfoo, and its cookbook guide to moving from Evals to Promptfoo walks through recreating each eval as a promptfooconfig.yaml with your prompts, model, test cases and assertions. You run it with promptfoo eval -c promptfooconfig.yaml --no-cache and view results locally with promptfoo view. Two warnings from that guide are worth taking seriously: similarity-based scores won’t match your old numbers exactly, and any recreated LLM-as-judge grader should be validated before you use it to decide whether a change ships.
Save your eval datasets first, even if you haven’t chosen a tool. A few dozen inputs with expected outputs are what tell you the migrated agent behaves like the old one.
Agents SDK or Workspace Agents: how to decide
OpenAI offers two destinations. The migration guide calls the Agents SDK “best for building agents through code” and Workspace Agents “best for building agents through natural language and sharing them with teams.” Workspace Agents need a ChatGPT Business, Enterprise or Edu workspace, and OpenAI warns that “workflows with strong determinism at their core may not migrate faithfully.”
| Your workflow | Move it to |
|---|---|
| Customer-facing bot behind ChatKit on your site | Agents SDK plus a self-hosted ChatKit server |
| Runs inside your product, triggered by your app | Agents SDK |
| Fixed steps, approvals or branching that must run the same way every time | Agents SDK |
| Internal helper your team uses in ChatGPT, mostly judgment and lookups | Workspace Agent |
| Nobody has used it in a month | Turn it off and save the export |
If you’re unsure, pick the SDK. Code you own outlives the next product change, and a workspace agent can always be added later for the internal version.
A plan for the 18 days left
Counting from November 12:
- By November 15: get everything out. Export code, screenshot canvases, list credentials, copy prompts, download eval datasets and results. Turn off workflows nobody uses.
- November 16-22: rebuild and wire up. Port each kept workflow to the SDK, reconnect file search with the same vector store IDs, recreate evals in Promptfoo and stand up the ChatKit server if you have one.
- November 23-27: run both side by side. Send the same eval set through the old workflow and the new code and compare answers, tool calls and refusals. Fix the gaps.
- November 28-29: switch traffic. Point ChatKit’s
apiURLat your server and remove prompt IDs from your code. Leave a day of margin before the shutdown.
If you have more than three or four workflows, you won’t rebuild all of them in this window. Rank by who would notice if it stopped, and move those first.
FormulaBot’s 1.5M+ users run on coded agents
The pattern this migration pushes you toward is the one FormulaBot (now Better Analyst) has run on as it grew to 1.5M+ users. For its analysis agents, our engineers built an LLM-driven code execution engine with LangChain agents and LangGraph workflows. A FastAPI backend streams results over WebSockets, and the Python the agents write runs in E2B sandboxes, so users can watch an analysis run and debug it as it happens.
None of that lives in a hosted canvas. The agent logic, the prompt engineering pipeline and the validation layer are in the codebase, versioned and reviewed like any other code. When a vendor changes a product, the team changes a dependency rather than rebuilding the product. That’s also how we build custom AI agents for clients today, and the other agent and LLM work we do is collected on our AI services hub.
If you’d rather have someone check the plan before you start, an AI audit maps each workflow to its new home, flags the ones that need a full rebuild, and ranks them by risk.
Frequently asked questions
- When does OpenAI Agent Builder shut down?
- Agent Builder is scheduled to shut down on November 30, 2026. OpenAI announced the deprecation on June 3, 2026. The hosted Evals dashboard and API and the v1/prompts API for reusable prompt objects are scheduled to shut down on the same day, and existing evals became read-only on October 31, 2026. ChatKit, the embeddable chat interface, remains available.
- Does the Agent Builder code export work without changes?
- No. OpenAI's migration guide says the export does not convert your workflow graph or guarantee that every behavior transfers unchanged. You get Agents SDK code in TypeScript or Python as a starting point. Plan to wire up tools, credentials and storage yourself, then test the exported agent against the same inputs the original workflow handled before you switch traffic.
- Will my ChatKit chat widget keep working after Agent Builder shuts down?
- ChatKit itself stays, but the hosted setup points the widget at an Agent Builder workflow ID, and that backend is going away. OpenAI's ChatKit docs tell teams migrating before the shutdown to use the advanced integration: run your own ChatKit server with the openai-chatkit Python package, store threads in your own database and point the widget's apiURL at your endpoint.
- Do I need to re-upload my file search documents?
- Usually not. The Agent Builder file search node reads from OpenAI vector stores by ID, and the Agents SDK FileSearchTool takes the same vector store IDs. The Vector Stores API is not on OpenAI's deprecations list as of November 12, 2026. Note each store ID from your workflow and pass it to the tool in your new code.
- Should I move to the Agents SDK or ChatGPT Workspace Agents?
- Use the Agents SDK if customers use the agent, if it runs inside your product, or if it needs fixed steps. Use Workspace Agents for tools only your own team uses inside ChatGPT, which needs a Business, Enterprise or Edu workspace. OpenAI warns that workflows with strong determinism at their core may not migrate faithfully to a workspace agent.