OpenAI is shutting down Agent Builder. What agencies do now
OpenAI retires Agent Builder on November 30, about fourteen months after it launched. Here is what else shuts down with it, what breaks on the client's end when a layer is retired, and how to tell in advance which layers are safe to put a paying client on.
Lumina··13 min read
OpenAI is shutting Agent Builder down on November 30, 2026. If a client system of yours runs through one of those visual workflows, you have about two weeks to move it into the Agents SDK and run it on your own infrastructure, or rebuild it somewhere else. The deprecation notice went out on June 3, 2026, and the OpenAI deprecations page lists the replacement as "Agents SDK or ChatGPT Workspace Agents."
That is the urgent answer. The useful answer is underneath it. Agent Builder launched on October 6, 2025. It was marked for shutdown eight months later. Nothing went wrong, nobody was punished, and the product was not a failure. It was a beta that stopped being the direction. If you run a small AI agency, that is the part worth sitting with, because the same thing will happen to a layer under one of your clients, and probably on a shorter clock than you are planning for.
AI agency software for sales and fulfillment
Custom fulfillment from our engineering team, live AI demos, and a client portal, all 100% white-labeled.
This post covers what is actually being retired, what breaks on the client's end when it happens, how to tell in advance which layers are safe to put a paying client on, and what to do in the next two weeks if you are one of the people with a live workflow on the canvas.
Is Agent Builder actually shutting down?
Yes. The Agent Builder guide now opens with the line "OpenAI is deprecating Agent Builder" and gives the shutdown date as November 30, 2026. Existing users can keep working during the transition window. After that date the canvas goes away.
Agent Builder was the visual piece of AgentKit, announced on October 6, 2025 alongside ChatKit, the Connector Registry, and the new Evals features. It was a drag-and-drop canvas in the playground where you wired nodes together, previewed runs, published versioned snapshots, and graded traces. It shipped in beta and stayed in beta. ChatKit went generally available and ChatKit is the piece that survives.
The replacement path is code. You download the workflow as an Agents SDK project and run it yourself, or you move the work into ChatGPT Workspace Agents. Neither of those is a one-click migration for a non-technical agency owner, which is the whole problem with having put a client on the canvas in the first place.
What else goes dark on the same day
November 30 is a bigger date than the Agent Builder headline suggests. Two other entries on the deprecations page share it, and both will bite anyone who built client systems inside OpenAI's dashboard rather than inside their own code.
Reusable prompt objects and the /v1/prompts endpoint are being retired. The recommended fix is blunt: move prompt content into application code. If you have been storing a client's receptionist prompt as a saved prompt object and referencing it by ID, that prompt is configuration living in a vendor's dashboard, and it is leaving.
The Evals dashboard, API, and graders go read-only on the same date, with Promptfoo named as the replacement. If your quality process for a client build was a saved eval suite in the OpenAI console, that process needs a new home.
And the Assistants API, which a lot of early agency builds ran on, was already shut down on August 26, 2026, with the Responses and Conversations APIs as its replacement. Then on December 11, 2026, gpt-5, gpt-5-mini, and gpt-5-nano retire too, pointing at gpt-6.1-sol, gpt-6.1-terra, and gpt-6-luna. We wrote about what that last swap does to chatbot costs in what GPT-6 Luna means for AI agency chatbot costs.
Read that list as one list and the pattern is obvious. Over roughly four months, a small agency running client systems directly on OpenAI gets a retired agent framework, a retired prompt store, a retired eval tool, and a retired model family. Each one has a reasonable replacement. Each one is also an evening of your time per client, and a conversation with a client who did not know any of this was coming.
What actually breaks on the client's end
This is where agency owners underestimate the cost, because the technical work is usually small and the client work is not.
A retired model ID fails hard. Anthropic's own model deprecations page says it plainly: requests to a model after its retirement date fail. There is no graceful fallback to a newer model. The receptionist that answered the phone yesterday returns an error today, and the client finds out the way clients always find out, which is a missed call from a customer.
Configuration that lived in the vendor's console disappears with the console. Code you kept is still code you have. A prompt you typed into a dashboard field is a thing you have to remember to export.
Phone call flows are the worst case, because the failure is silent to you and loud to them. Twilio gives a clean historical example. In a changelog entry dated November 27, 2023, Twilio announced that support for the Google Dialogflow ES integration would end on May 27, 2024, about six months later, and said directly that call flows connected to ES bots through Twilio would stop working and cause traffic interruptions. Six months of notice is generous. It still means every agency with a client on that integration had a migration project they did not quote for.
Then there is the part nobody budgets for. You have to tell the client. If you sold a custom AI system and then send an email in month seven that says we need to migrate your system because our vendor is retiring a component, you have handed a nervous buyer a reason to ask what else is temporary. Some of them will ask whether they should just wait for this stuff to settle down. A few will cancel. That is the real cost of platform risk, and it does not show up in the migration estimate.
Which vendors tell you in advance, and which do not
Here is the comparison that actually matters when you pick what to build on. Not features. Not price per minute. Whether the vendor has written down what happens when they take something away.
Anthropic publishes a notice floor. The deprecations page states that it notifies customers with active deployments and provides "at least 60 days' notice before model retirement for publicly released models." You can argue 60 days is short for an agency with twenty clients. You cannot argue about what the commitment is, because it is in the documentation, with a dated table of every retirement underneath it. Claude Sonnet 4.5 is in that table: announced September 30, 2026, retiring November 30, 2026.
OpenAI publishes a deprecations page too, with notice dates and shutdown dates for every model, endpoint, and feature. Agent Builder got about six months. The Assistants API got longer. The page is public, dated, and specific, which is the thing you want.
Now check the vendor you are actually reselling. Most of the voice AI platforms and agency-in-a-box products in this market publish no deprecation policy at all. No notice floor, no dated retirement table, no public commitment about how long a feature stays. That is not an accusation of bad faith. Small companies move fast and do not want to write promises they might break. But it means your real notice period is however long it takes you to notice something changed.
So the uncomfortable ranking: the large model vendors are the most predictable layer in your stack, and the small white-label tools wrapped around them, including ours, are the least documented. If you want that argument in full, we went through it in build your own AI stack, or use a white-label one.
How to tell whether a layer is safe to put a client on
Four questions, in order of how much they tell you.
Is it generally available or is it beta? Agent Builder was beta for its whole life and ChatKit was GA. One is gone and one remains. Beta means the vendor is still deciding, and a vendor still deciding is not a place to park a client's phone line.
Is the thing you would lose code or configuration? If your work lives as code in a repository you control, a vendor retirement is a dependency bump. If it lives as fields in someone's web form, a vendor retirement is a rebuild from memory. This is the single most useful test, and it applies to visual builders of every brand, including n8n, which has its own license and hosting questions we covered in can you run an AI agency on n8n.
Can you get your work out? Not theoretically. Try it. Export one client's configuration today and see what the file actually contains. If the export is a screenshot, you do not have an export.
Is there a dated, public deprecation page? If yes, read it before you sign the next client, not after. If no, assume your notice period is zero and price accordingly.
You will notice none of those questions are about how good the product is. A better product with no exit is a worse bet than a plain one you can leave.
What to do this week if you are on Agent Builder
A short sequence, in this order.
Make a list of every client system that touches a published Agent Builder workflow. Include your own internal tools, because the demo that wins your deals is also a dependency.
Screenshot and export each workflow now, before the shutdown window closes. Even an imperfect record of the node graph is faster to rebuild from than memory.
Pull the prompt text out of every saved prompt object and put it in a file you own. This takes twenty minutes and it is the step people skip.
Pick the replacement per client, not globally. A simple chat flow may move to ChatKit with the logic in your own code. A multi-step workflow probably wants the Agents SDK on your own hosting. A client system that is mostly a phone line may be the one where you stop assembling layers yourself.
Then send the client email before November 30, not after. Short, plain, no apology, no mention of the vendor's name if you do not want it: we are moving your system onto updated infrastructure this month, you will not see a change, here is the date. Clients forgive maintenance. They do not forgive surprises. We have a longer version of that conversation in who maintains the AI system after it goes live.
Where our own answer helps, and where it does not
The pitch writes itself here, so let me be careful with it.
What the Lumina Sales platform does about this problem is absorb the model and infrastructure layer. The voice and chat agents run on our infrastructure. There is no API key for you to obtain and no phone number for you to buy, which also means there is no model ID in your possession to go stale on December 11. When a model family retires, that is our migration to run, not yours, and your client does not get an email about it. The same is true of the demo site you send prospects. You do not maintain it.
That is a real benefit and it is also a trade, so here is the honest side of it.
Using a white-label platform does not remove platform risk. It moves the risk to a smaller company. We are a small company. We do not publish a dated deprecation table with a 60-day notice floor, and until we do, you should not treat us as more durable than OpenAI or Anthropic just because we are friendlier to deal with. That comparison runs the other way.
Our export story is honest but limited. If you cancel, nothing is deleted. Published demos unpublish so the links stop serving, and everything you built stays as a draft you can bring back by resubscribing. That protects you from losing work. It is not the same as a file you can take to a different vendor, and I am not going to pretend it is.
And if you have genuine engineering capacity, owning the code is better. Someone who can run an Agents SDK project on their own hosting, read a deprecations page, and bump a model ID in an afternoon should do exactly that. They will have more control than we can give them, and the honest reason to use something like ours is that most agency owners selling five-figure systems to roofers cannot do that and should not try to learn it on a client's phone line. If that is not you, our pricing page is not going to change your mind, and it should not.
Who this is wrong for: anyone whose client has procurement. If your buyer's legal team wants a contractual commitment about platform lifecycle, uptime, and notice periods, a $49 a month tool is not where that comes from, ours included. Go direct to a vendor who will sign that, or write the commitment yourself and carry it.
What to put in the client contract
Two clauses handle most of this, and neither one is adversarial.
A substitution clause. State that the system may be moved to equivalent or better underlying technology as providers change, at no cost to the client, with notice but without renegotiation. Without it, every vendor retirement is technically a change to what you sold.
A maintenance line that names who absorbs this work. If your monthly fee covers keeping the system running through provider changes, say so, because that sentence is the justification for the monthly fee. If it does not, say that too and price migrations separately. The ambiguity is what causes the fight. We have a fuller list in what to put in an AI services contract.
FAQ
Is my client's system broken right now if it uses Agent Builder?
No. The shutdown date is November 30, 2026, and existing users keep working until then. After that date the canvas is gone, so a workflow you have not moved stops being something you can edit or rely on.
Does this mean OpenAI is unreliable to build on?
The opposite, if you read the whole page. OpenAI and Anthropic both publish dated deprecation tables with announcement dates, shutdown dates, and named replacements. That is more notice than almost anything else in this market gives you. The mistake was not using OpenAI. It was putting client configuration in a beta canvas instead of in code.
How often will I have to do a migration like this?
Based on the public deprecations pages, plan on a few per year touching something in your stack, most of them small. The OpenAI page alone has four separate shutdown dates between August and December 2026. Budget an afternoon per client per year and you will mostly be right.
Should I tell clients which vendors are under their system?
I would not volunteer the list, and I would never lie about it if asked. Naming your vendors invites the client to go direct and makes every vendor's news your news. Refusing to answer a direct question costs you the deal. Say that you manage the underlying providers and that keeping the system current is part of what they pay for.
What if I built my own stack instead?
Then you own both the control and the work. You read the deprecations pages, you bump the model IDs, you run the migrations, and nobody else's roadmap decides your week. For a technical founder that is the right call.
The next step
Open the OpenAI deprecations page and the Anthropic deprecations page right now and search both for every model and endpoint name that appears in your client systems. It takes fifteen minutes and it will tell you what your next three months look like.
If the honest answer is that you do not know which model IDs your client systems use, that is the finding, and it is worth more than this post. Either learn the stack properly or stop owning that layer. If you want to stop owning it, talk to the live demo on our home page, read the FAQ including the parts about caps and cancellation, and then start an agency workspace for $49 a month and let the model migrations be our problem.