What Is Shutting Down
Agent Builder, Evals, and the v1/prompts endpoint. The shutdown is product-wide with no regional exceptions.
The deprecation covers three related features. Agent Builder is the visual workflow editor. Evals is the testing framework that let you measure agent performance. The v1/prompts endpoint stored reusable prompts. All three shut down on November 30, 2026.
ChatKit, the runtime that actually executes agents, remains available. But the convenient setup where ChatKit talked to an OpenAI-hosted backend that was your Agent Builder workflow: that is what ends. After November 30, there is no Agent Builder backend to call.
OpenAI has not stated whether workflows published through Agent Builder will continue to execute via ChatKit after the deadline. Multiple analyses conclude that existing live embeds should be treated as at risk. If your agent is running in production, assume it stops working on December 1 unless you have migrated.
What You Can Export
OpenAI provides a code export that converts your agent to TypeScript or Python using the Agents SDK. This sounds like portability. It is not.
The code export gives you a starting point. OpenAI's migration guide explicitly states that it does not convert your workflow graph or guarantee identical behavior. The visual structure you designed, the thing that made Agent Builder no-code in the first place, is not represented in the export. You get code that approximates what the agent did, which you now have to read, understand, debug, and maintain.
Knowledge bases may also need rebuilding. Agent Builder's file search uses vector embeddings tied to the platform. Whether those embeddings transfer, or whether you have to re-index your documents, depends on implementation details OpenAI has not fully documented.
OAuth connections are another issue. Every app connection was authorized against the old platform's client ID and redirect URI. After migration, those connections may need re-authorization. If your agent talks to external services, plan for users to re-authenticate.
| Asset | Transfers? | Notes |
|---|---|---|
| Code export | Partial | Starting point only; does not match workflow graph |
| Workflow graph | No | Not represented in export |
| Knowledge base embeddings | Unclear | May need re-indexing |
| OAuth connections | No | Users may need to re-authorize |
| Prompt definitions | Partial | Copy them manually |
| Eval results | No | Download before shutdown |
The No-Code Lock-In Pattern
No-code platforms trade learning curve for dependency. The trade is invisible until you need to leave.
The value proposition of Agent Builder was clear: build AI agents without writing code. Drag components, connect them visually, deploy. For teams without developer resources, or for prototyping ideas quickly, this was genuinely useful. The product worked.
The hidden cost was that the visual workflow existed only inside the platform. There is no standard format for AI agent workflows. There is no interoperability between Agent Builder and competing tools. The ease of building was real; the inability to export was also real, just less visible until now.
This is not unique to OpenAI. It is the pattern of no-code platforms in general. The same tradeoff appears in automation tools, database builders, and website constructors. The platform handles complexity for you, but it handles it in a proprietary way. When the platform changes direction, or raises prices, or shuts down, your investment is stranded.
One commentator called the Agent Builder shutdown the biggest validation no-code platforms will ever get. The argument is that dedicated no-code tools, the ones whose whole business is the no-code experience, are less likely to shut down their core product than an AI company adding no-code as a feature. Maybe. The deeper lesson is that any workflow you cannot export is a workflow you do not own.
Migration Paths
OpenAI lists two official destinations: the Agents SDK and ChatGPT Workspace Agents. Each has tradeoffs.
Agents SDK
The SDK is code. Python or TypeScript. If you have developers, this is the path that gives you the most control. The code export provides a starting point, which you then modify and extend. You can host it yourself, version control it, and understand exactly what it does. The cost is that you now have code to maintain, which was the thing no-code was supposed to avoid.
ChatGPT Workspace Agents
Workspace Agents is a different product, aimed at enterprise teams using ChatGPT. It is not a direct replacement for Agent Builder. The capabilities may or may not match what you built. Whether your workflow translates depends on what you were doing and how Workspace Agents works, which is still evolving.
A third path is to rebuild in a different tool entirely. Dedicated agent-building platforms exist, some from companies whose entire business is that platform rather than a side feature. The rebuild cost is real, but so is the reduced risk of another shutdown. The question is whether you trust the next platform more than you trusted this one.
What This Teaches About AI Platform Risk
Agent Builder joins a growing list of AI platform features that shipped, attracted users, and then shut down. The pattern is not malice; it is economics. AI companies are burning capital to find product-market fit, and features that do not contribute to that fit get cut. Enterprise products subsidize consumer products and then get deprecated. The timeline from launch to sunset can be under two years.
The practical lesson is to treat any workflow you build on an AI platform as temporary. Export what you can, when you can. Document what the workflow does in terms that do not depend on the platform. Keep your source materials, documents and prompts and configurations, in files you control. The platform is a tool for building; the files are what you keep.
This is not an argument against using AI platforms. They provide capabilities you cannot easily build yourself. It is an argument for using them with eyes open. The convenience is real. So is the dependency. When the platform moves, you move with it or you rebuild. There is no third option where your investment is protected by goodwill.
Keep your actual knowledge in a system you control, even when AI platforms help you work with it. Portable formats →
What to Do Before November 30
- Export your agents now. Use the code export even if you do not plan to use the code. It is documentation of what you built.
- Download your Eval results. They are not migrating. If the data matters, save it.
- Copy your prompt definitions. The v1/prompts endpoint is shutting down. Put them in files you control.
- Document your workflows in plain language. Describe what each agent does, what tools it uses, what the expected behavior is. This is what you will use to rebuild if the code export is not sufficient.
- Test the migration path you choose. If you are going to the Agents SDK, run the exported code and compare behavior. If you are going to Workspace Agents, verify that your use case is supported.
- Plan for re-authorization. If your agents connect to external services, users may need to re-authenticate after migration. Communicate this before the deadline.
- Consider whether to rebuild on a different platform. The shutdown is a forcing function. If you were unhappy with Agent Builder anyway, this is the moment to evaluate alternatives.