OpenAI Agent Builder Is Shutting Down. Here Is How to Move Your Workflows, Free.

If you built workflows in OpenAI’s Agent Builder, they now have an end date. OpenAI is shutting Agent Builder down on November 30, 2026. Nothing you built was wrong. OpenAI is moving people to its Agents SDK and to ChatGPT workspace agents, and the visual editor is going away.

Five weeks is enough to move, as long as you start now. This guide covers what you can take out of Agent Builder, where each piece lands in FlowRunner, why the approval step is the one to move first, and what changes about the bill.

The short version, for anyone who got here from a search box. Agent Builder shuts down on November 30, 2026. Its export gives you each workflow as Agents SDK code, and FlowRunner Prompt to Flow builds a FlowRunner flow from that code. FlowRunner’s Free plan has no card and no time limit, counts one execution per complete run, runs your agents on your own OpenAI key, and includes every integration and human-in-the-loop approvals, with 14 days of Professional on signup.

To skip rebuilding by hand, follow how to convert Agent Builder workflows to FlowRunner. It walks through the five steps and six of OpenAI’s own templates converted this way.

What’s happening, and when

OpenAI’s deprecations page lists the dates:

• Agent Builder is scheduled to shut down on November 30, 2026. ChatKit stays available.
• Evals: existing evals become read-only on October 31, 2026, and the Evals dashboard and API shut down on November 30. If your workflows have evals, save the results before then.

OpenAI’s migration guide offers two paths. You can run the exported code in your own application with the Agents SDK, which means hosting and maintaining it yourself. Or you can create a ChatGPT workspace agent from the export, which needs a ChatGPT Business, Enterprise, or Edu workspace. OpenAI’s guide also says the export doesn’t convert your workflow graph or guarantee that every behavior transfers unchanged, so plan to review and test whatever you move.

What the export gives you

Agent Builder has one export. Open a workflow, click Code, choose the Agents SDK tab, and pick TypeScript or Python. You get the workflow as code: each agent’s instructions, model, and settings, the structured outputs, the branches, and the order everything runs in.

What’s missing matters as much. We exported six of OpenAI’s own templates, and here’s what we found:

• Web Search breaks the export. If any agent has the built-in Web Search tool, Agent Builder refuses with “Unknown tool type: web_search”. You have to remove the tool to get the code.
• Human approval comes out as a stub. The approval step became a function that always returns true.
• Function tools come out empty. A custom tool’s body was a TODO.
• Uploaded files aren’t in it. The exported input is the chat text only.
• Some exports don’t run. One referenced a variable it never defined, and one guardrail’s input was a 32,000-character escaped string.
• Your data stays at OpenAI. File search points at vector stores in your OpenAI account, and the export carries only their IDs.

Copy the code for every workflow now, even the ones you plan to retire. Once Agent Builder is gone, you can’t open them to get it.

What to do with the code

You don’t rebuild from it by hand. Even where the code is incomplete, it still says what each workflow does, and FlowRunner Prompt to Flow reads it that way. It builds the flow and asks you about whatever the code leaves out. For each workflow:

1. Connect FlowRunner Prompt to Flow to your AI tool. It’s an MCP server you add to Claude, ChatGPT, Claude Code, Codex, or Cursor, at https://app.flowrunner.ai/mcp. You sign in with your FlowRunner account, and the Free plan needs no card. You only do this once.
2. Paste the code with a short request. Start a new chat, ask it to rebuild the Agent Builder workflow as a FlowRunner flow, and paste one workflow’s code. Mention what the export dropped, such as which agents had Web Search and who should approve requests.
3. Answer its questions. It tells you what it can’t map one to one and asks about anything the code doesn’t say. Then it builds the flow in your workspace.
4. Add your OpenAI key and connect your accounts. Your agents run on your own key, and approvals need an email or Slack account. Each is set up once and shared by all your flows.
5. Test, then switch over. Run the flow with the messages you used in Agent Builder’s Preview, compare the answers, and point your app at the flow.

The conversion guide has the exact request to paste, the setup steps for each AI tool, and six of OpenAI’s own templates converted this way.

Where each Agent Builder piece lands

This is the table to keep open.

Agent Builder	FlowRunner	What changes
Workflow	Flow	Starts from an API call, a webhook, a schedule, an incoming email, or another flow, not only a chat.
Start, with input_as_text	The flow’s Initial Data	You can name as many input fields as the flow needs.
Agent	AI Agent	Same instructions and model, on your own OpenAI key. You can switch any agent to another provider later.
Structured output	The agent’s parsed output	Later steps read each field by name.
Guardrails (moderation, jailbreak, PII)	A Condition with an AI Question	A small model answers yes or no about the input, and the flow takes the matching path.
If/else	Condition, Value Router, AI Router	Value Router for several fixed paths; AI Router branches on meaning.
Human approval	A human-in-the-loop step	The person is reached over email, Slack, WhatsApp, phone, or a Form, and the run waits for the answer.
Web search	A web search tool attached to the agent	Tavily, Perplexity, Firecrawl, You.com, and others are available as agent tools.
File search	A Knowledge Base	Upload your documents again; the agent searches them as a tool.
MCP	MCP Extensions on the agent	Attach a whole server or pick individual tools.
Function tools	An integration action, a Custom Cloud Code step, or another flow the agent calls	Whatever the stub was supposed to do, built for real.
Code interpreter	No direct equivalent	Custom Cloud Code runs JavaScript that you write ahead of time.
While	Repeat
Transform	Transform Data	40+ common operations without code.
Set state	Set Variables, or Shared Memory for state that lasts across runs
End	Return Result	Sends the answer back to whoever started the run.
Preview and traces	Run history	Every run is recorded block by block, with each block’s input and output.

One change touches every row. An Agent Builder workflow runs inside a chat, so the person is always right there. A FlowRunner flow runs on its own, started by a call or an event, and it contacts a person only when a step needs one. If your workflow asked the user a follow-up question in the chat, decide where that question goes now: back to the app that called the flow, or out by email or Slack.

The approval step is the one to move first

Agent Builder’s human approval node asks the person in the chat to approve or reject, and the workflow continues down the matching path. In the code export it doesn’t ask anyone, because it comes out as a function that always approves.

FlowRunner treats the person as part of the run. Human-in-the-loop is an execution pattern where AI agents pause autonomously, assemble the relevant context and the decision choices available, route to a human via their preferred channel, and resume the moment the human responds. The run reaches the person over email, Slack, WhatsApp, phone, or a Form. It waits up to 30 days on Free, Starter, and Growth and up to a year on Professional and above, and then picks up where it left off with the answer.

If one of your workflows has an approval in it, move that one first. It’s the one the export breaks, and it’s the one that tells you whether the new platform fits.

What changes about the bill

Agent Builder had no fee of its own. OpenAI billed you for the model tokens your agents used and for its built-in tools.

On FlowRunner, the OpenAI part stays the same. FlowRunner uses BYOK, a model where you supply your own AI provider API credentials, pay the providers directly for usage, and FlowRunner orchestrates calls across providers without taking a markup on inference. Your agents call OpenAI with your key, at your rates, on your invoice from OpenAI.

What FlowRunner counts is the execution. An execution is one complete run of a flow, start to finish. A run with eight agents counts once, the same as a run with one, and a run that waits two days for an approval still counts once.

Where this lands

The Free plan is $0, permanent, no card, and no time limit. It runs 100 executions a month, one at a time, with 24 hours of log history. Unlimited users, unlimited workflows, and every integration, with AI agents and human-in-the-loop included. One Free workspace per account.

Every new account also gets 14 days of the Professional plan on its first workspace, with audit trails and role-based access control, and no card required. When the trial ends without a payment method, the workspace moves to Free and the flows keep running. Above Free, Starter is $5 a month for 300 executions or $15 for 3,000, and the ladder continues from there.

Pipedream Workflows users are going through the same thing on a longer clock; their guide covers it from their side.

Start on the Free plan

Open a free workspace, keep the exported code on the other screen, and move the workflow with the approval in it first. No card, no clock. If you’d rather have someone go through your workflows with you, send us a note. Migration help for people coming from a shutdown is free.