Agent Builder Shutdown: You Built a Workflow, Not an Agent

David AlfordOctober 14, 20269 min read

The OpenAI Agent Builder shutdown is set for November 30, 2026[1]. OpenAI’s migration guide offers two ways out, and its own fine print admits a problem with one of them: “Workflows with strong determinism at their core may not migrate faithfully to a workspace agent.”[2]

That sentence describes most of what people built on the canvas. You drew fixed steps with a model call at a few of them. One exit keeps those steps and asks you to write and host code. The other keeps the no-code and swaps your steps for a prompt. Neither keeps both, and if a client relies on what you built, that’s the thing to sort out before the date.

When does OpenAI Agent Builder shut down?

Agent Builder shuts down on November 30, 2026. OpenAI announced the deprecation on June 3, about eight months after launching the product on October 6, 2025. The Evals platform and reusable prompt objects end the same day, and existing evals become read-only a month earlier, on October 31[1][3].

October 31, 2026. Existing evals become read-only.
November 30, 2026. Agent Builder shuts down. So do the Evals dashboard and API, and the v1/prompts API with its reusable prompt objects.
Staying. ChatKit, the embeddable chat window. OpenAI’s notice says “ChatKit remains available.”[1]

That last line needs a second read. ChatKit is only the chat window. If yours is connected to an Agent Builder workflow, OpenAI’s docs say you can keep using that hosted workflow “during the Agent Builder transition window”, and that anyone migrating should connect ChatKit to their “own server-side agent implementation”[4]. OpenAI hasn’t said exactly what such a widget does on December 1. I’d assume it stops answering.

OpenAI hasn’t given a reason either. The deprecations page, the changelog and the migration guide don’t offer one.

You built a workflow, not an agent

Most Agent Builder projects are workflows: a fixed set of steps in a fixed order, with a model doing one job at some of them. An agent is different, because it decides for itself what to do next. The product was called Agent Builder, and what it mostly produced was the first kind.

Engineers call the first kind deterministic: the same input takes the same route through the steps every time. The model’s wording can still vary from run to run. The route can’t.

Anthropic drew the line this way in 2024. Workflows are “systems where LLMs and tools are orchestrated through predefined code paths.” Agents are “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”[5]

Now look at what the canvas gave you. OpenAI’s node reference lists Start, Agent, Note, File search, Guardrails, MCP, If/else, While, Human approval, Transform and Set state[6]. Most of that list is plumbing for a predefined path: branches, loops, approvals and data handling. The example OpenAI gives for If/else is a model sorting a question into a category and the flow routing it to the matching agent.

OpenAI’s own guide to building agents tells readers to check whether they need one at all: “Otherwise, a deterministic solution may suffice.”[7] For a support triage flow or a lead-qualification bot, it usually does.

Each exit keeps half of what you had

OpenAI names two destinations. The Agents SDK keeps your fixed steps and takes away the no-code and the hosting. ChatGPT Workspace Agents keep the no-code and take away the fixed steps. The bridge to either one is a block of exported code you copy by hand, one workflow at a time[2].

The Agents SDK: your steps, your servers

OpenAI describes the SDK as “Best for building agents through code.” You open a workflow, choose Code, choose Agents SDK, pick TypeScript or Python and copy what appears. The guide is plain about what that is: the process “does not convert your workflow graph or guarantee that every behavior transfers unchanged.”[2]

From there the branching is code someone has to maintain, and the thing that used to run on OpenAI’s side needs a server, credentials and deploys. One developer on OpenAI’s forum, who ran Agent Builder with ChatKit “for multiple clients”, wrote that managing the backend himself “makes the use of this tool economically unviable”[8].

On October 1, an OpenAI support reply on the same forum recommended a third option for production workflows, the Agents API[9]. It’s an API, so you still need someone who writes code against it, and the migration guide doesn’t mention it yet.

Workspace Agents: no code, no fixed path

The other exit is “Best for building agents through natural language and sharing them with teams.” It needs a ChatGPT Business, Enterprise or Edu workspace. The conversion is a chat message: you paste the exported code after the line “Please help me convert this workflow into an agent:” and review what comes back[2]. This is the path the determinism warning is attached to.

It’s also an awkward fit for anything a client’s customers talk to. A workspace agent can be started from your own software, but OpenAI’s docs say the call returns a link to a ChatGPT conversation and that “The agent’s response cannot currently be retrieved through the API.”[10] My read is that it suits work your own team does inside ChatGPT, and that open-ended jobs like research or drafting are where it earns its place.

What you can still show a client after November 30

Three things you could point to in Agent Builder go away with it: the drawn graph that showed every branch, a published version you could pin by ID, and the hosted Evals dashboard where you graded behavior. Each exit replaces them with something you assemble yourself, or with a prompt.

The graph. In Agent Builder, “Each connection between nodes becomes a typed edge”[11]. A client or a colleague could read the flow. In the SDK it’s code, and in a workspace agent it’s written instructions.
Pinned versions. Publishing created “a new major version that acts as a snapshot”, and you could “specify an older version in your API calls”[11]. With the SDK, that job moves to your own repository and deploys.
Evals. The dashboard and API end on November 30. OpenAI points to Promptfoo, an open-source tool you run yourself[12].
Traces. OpenAI hasn’t said whether traces from existing workflows stay viewable after the shutdown. Save what you need now.

I’m not claiming a workspace agent keeps no records. The loss is narrower than that, and it’s the one clients care about. When someone asks in January why the bot told a customer something, you want to say which version was live and what each step received and returned. That’s all an audit trail is. Ask your new home how it answers that before you move.

How to keep fixed steps without writing code

Rebuild the flow in a workflow tool, where the steps are drawn on a canvas and a model is one step among them. That’s the category Agent Builder was really in. I build one of these, TaskJuice, so here’s what ours does with a flow like yours and where it falls short.

Branches run on data. A published workflow follows the steps you drew. Branch and Switch conditions are checked against the data, and a model only picks a path where you place an AI Classify or AI Agent step.
Versions are pinned. A run finishes on the version it started on. Editing a live workflow creates a draft that changes nothing until you publish, and you can switch a previous version back on.
Agents are fenced in. An AI Agent step can only call the tools you wire to it, stops after a set number of turns (10 by default, 50 at most), and a tool that takes an action can be set to wait for a person’s approval.
Every run leaves a record. You can see which version ran, what triggered it, and each step’s input, output, timing and error. Approvals record who decided and when.
Chat on a client’s site. A chatbot answered by a workflow, added to a page with one line of code, is our nearest thing to Agent Builder plus ChatKit.

You don’t have to redraw anything by hand. Copy the Agents SDK export, connect Claude or ChatGPT to the TaskJuice MCP server, paste the code and ask for the same flow as a draft. The assistant tells you what it couldn’t match, and you connect the accounts yourself. The MCP setup guide has the steps.

Where TaskJuice falls short

AI steps still vary. Fixed steps don’t make a model’s wording repeatable. Setting the temperature and requiring typed fields from an extraction step narrows it. Nothing removes it.
Records expire. Run history is kept 3 days on Free and Solo, 7 on Starter, 30 on Growth and 90 on Scale. There’s no export of run history and no account-wide log of who changed what.
Evals are advice. You can score an AI step against a dataset, but a failed check doesn’t stop a publish.
No export file and no certifications. A workflow can’t be downloaded as a file today, and we hold no compliance certifications.
Chat is plainer than ChatKit. Visitors can’t upload files, and replies only stream word by word in the simplest setup.

There’s more on where an agent is the wrong tool in our comparison of workflows and AI agents.

What to do before October 31 and November 30

Work backward from two dates. By October 31, export any eval results you want to keep, because evals go read-only that day. By November 30, export every workflow’s code, list every place a workflow is in use, and have a replacement answering before you switch anything off.

This week. List every workflow and where it runs: which client, which website, which widget. The ones nobody remembers are the ones that go dark.
Before October 31. Export eval results and the datasets behind them.
Early November. Export each workflow’s code even if you’ll never run the SDK. It’s the only description of the flow you can take with you. Do it early, because users on OpenAI’s forum reported exports failing for workflows that use the web search tool[13].
Mid November. Rebuild, then run old and new side by side on real questions and compare the answers.
Before November 30. Swap the widget and tell the client. Copy out any saved prompts too, since reusable prompt objects end the same day and OpenAI’s advice is to “move reusable prompt content into your application code”[1].

Frequently asked questions

Is ChatKit shutting down too?

No. OpenAI says ChatKit remains available. What ends is the Agent Builder workflow behind a hosted ChatKit setup, so a widget connected to one needs a new backend before November 30.

Can I import an Agent Builder workflow into another tool?

Not directly. The only export is Agents SDK code in TypeScript or Python, copied one workflow at a time. Other tools can’t load it, though an AI assistant can read it and rebuild the steps.

Is all of AgentKit being deprecated?

No. The June 3 notice covers three things: Agent Builder, the Evals platform and reusable prompt objects. ChatKit and the Agents SDK continue.

What’s the difference between a workflow and an agent?

A workflow follows steps someone laid out in advance. An agent chooses its own steps as it goes. If you can draw what should happen on a whiteboard, you want a workflow, with a model at the steps that need one.

Agent Builder will have lasted under 14 months, and for the first eight nobody knew it was temporary. I don’t hold that against anyone who built on it. The canvas was good.

What I’d take from it is to be honest about what you built. If it’s a set of steps a client relies on, put it somewhere the steps are the product, and keep agents for work that really is open-ended. Relay.app’s customers got a harder version of this lesson in September, which I wrote up in Relay.app Shut Down: What Happened and What Users Lost.

[1] Deprecations, OpenAI API documentation: developers.openai.com/api/docs/deprecations

[2] Migrate from Agent Builder, OpenAI API documentation: developers.openai.com/api/docs/guides/agent-builder/migrate-from-agent-builder

[3] Changelog, OpenAI API documentation: developers.openai.com/api/docs/changelog

[4] ChatKit, OpenAI API documentation: developers.openai.com/api/docs/guides/chatkit

[5] Building effective agents, Anthropic: www.anthropic.com/engineering/building-effective-agents

[6] Node reference, OpenAI API documentation: developers.openai.com/api/docs/guides/node-reference

[7] A practical guide to building agents, OpenAI: cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf

[8] Deprecation notice: Agent Builder, OpenAI Developer Community: community.openai.com/t/deprecation-notice-agent-builder/1382650

[9] Agent Builder MCP Secrets API, OpenAI Developer Community: community.openai.com/t/agent-builder-mcp-secrets-api/1368113

[10] Trigger workspace agent runs, OpenAI Workspace Agents documentation: developers.openai.com/workspace-agents/trigger-runs

[11] Agent Builder, OpenAI API documentation: developers.openai.com/api/docs/guides/agent-builder

[12] Moving from OpenAI Evals to Promptfoo, OpenAI Cookbook: developers.openai.com/cookbook/examples/evaluation/moving-from-openai-evals-to-promptfoo

[13] Agent Builder code export, OpenAI Developer Community: community.openai.com/t/agent-builder-code-export/1384377