Let’s talk about the twenty-one days you have left to get your agent off their canvas — while the export button still works.
On the thirtieth of November, OpenAI shuts down Agent Builder. Not “stops developing.” Shuts down. The same day, the Evals platform goes dark, and so do reusable prompt objects — that whole v1/prompts surface where a lot of people keep the only copy of the thing that makes their agent work. It was announced back in June, so nobody can say they weren’t told. Twenty-one days is what’s left.
That’s one vendor. But it’s not one vendor.
Amazon closed Bedrock Agents to new customers on the thirtieth of July. Didn’t kill it — renamed it “Bedrock Agents Classic,” put it in maintenance, froze the model catalogue, and stopped letting anyone new in. The product launched in November 2023. It got three years. AWS’s own migration guide says two to four weeks per workload to get off it, and the import tool converts your configuration to LangGraph, which is a polite way of saying: here’s your settings, now go be a developer.
Google did the quieter version. Vertex AI Agent Builder stopped being its own thing and got folded into the Gemini Enterprise agent platform. Same floor, new building, new lease.
So in one year, three of the biggest companies on earth each retired the part of the product where a normal person could build an agent without writing code.
Here’s the story I keep thinking about. Not a specific person — a shape of person, and there are tens of thousands of them.
Somebody in operations, or a two-person consultancy, or the guy who runs IT and everything else at a 40-person business. In 2025 or early 2026 he opens a canvas. Drag a node. Connect a tool. Paste a prompt. Test it. It works. He builds fourteen of these things over a long weekend, and for the first time in his career he’s shipping software without a developer. He shows it to his boss. It saves the company a person’s worth of hours a week. He tells people at braais that this is the future and it’s free to try.
He is not being paranoid when he assumes it will still be there. Everything else he uses still works. His accounting package from 2014 still works.
Then a deprecation notice lands in June, and he reads it in August, and now he has nineteen days.
Everybody’s going to file this under “normal software lifecycle.” Products get retired. That’s the industry. Fine.
But look at what actually happens in the retirement, because it isn’t neutral.
The first thing: the export is not your agent. OpenAI’s own migration page says the export to the Agents SDK does not convert the workflow graph, and does not guarantee that every behaviour transfers unchanged. Read that again. They give you code — TypeScript or Python — that represents the pieces, and then tell you, in writing, that the thing those pieces made might not work the same way. So “export” isn’t a moving box. It’s a pile of parts and a note saying good luck.
The second thing: the people who used a canvas used it because they don’t read code. So the migration path for the least technical users is the most technical one available. There’s a second path — Workspace Agents in ChatGPT — and that’s the interesting one, because that path isn’t a migration. It’s a move. You stop owning a workflow and start describing one to a model inside somebody else’s product, where you can’t see the graph, can’t version it, and can’t take it with you next time.
Third thing, and this is the one that should make you sit up.
The Evals platform dies on the same day as the builder. Twenty-one days from now, a lot of companies lose the tool they used to build the agent and the tool they used to check whether the agent was any good, in the same afternoon. And OpenAI’s recommended destination for evals is Promptfoo. Which OpenAI acquired in March.
Now — that’s not a scandal. Promptfoo is a real product and a real team, and OpenAI buying it is a normal strategic move. But notice the shape of it. The company being tested now owns the most widely used independent testing tool. And the tool you were paying for, that lived inside their platform, is being switched off in favour of it. That’s not evil. That is what consolidation looks like from the inside, and you are standing inside it.
Which brings us to the pattern, because this is RADAR and the job is to name the game.
Platforms don’t retire the layer where you assemble things because it failed. They retire it because it succeeded at the only job it had: getting you in.
A canvas is a recruitment tool. It’s cheap to run, it needs no support contract, it produces a demo in an afternoon, and it converts a person who would never have called themselves a builder into somebody with fourteen workflows that his business now depends on. That’s a fantastic acquisition funnel. It is a terrible place to leave him, because a canvas holds no runtime, no memory, no meter, and no data gravity. You can leave a canvas. You can’t leave a runtime.
So the traffic all moves one direction. From the surface where you configured, to the surface where it runs, remembers your context, holds your tool integrations, logs your calls, and charges you per unit of thinking. AWS: AgentCore. Google: Agent Runtime with Memory Bank and an agent marketplace and a seven-hundred-and-fifty-million-dollar partner fund to get everyone building there. OpenAI: Agents SDK, or Workspace Agents inside ChatGPT. Anthropic: managed agents on its platform.
Different logos. Same direction. The vendor stops selling you the workbench and starts renting you the workshop.
And here’s the part people miss, the part that actually decides whether you’re fine or not.
The three-year lifecycle isn’t the problem. Fast-moving software has always been like that. The problem is that we let the fast-moving layer hold the slow-moving assets.
Your prompts, your test cases, your business rules, your knowledge base, the little exceptions that took you six months to discover and encode — those are the slow assets. They’re the actual value. They should live somewhere boring, in a format a human can read, in a repo you control, with a date on it. Instead they live in a prompt object that gets deleted on the thirtieth of November, and in an eval dashboard whose results nobody thought to copy out because who copies out a dashboard.
So let’s get practical, because there is still time and this is not a eulogy.
This week, list every agent or automation you have running on somebody’s canvas. Name the vendor, the date you built it, and the last time it actually ran. Anything that hasn’t run in ninety days, kill it. You don’t need to migrate a thing that isn’t working — that’s how you end up doing three weeks of unpaid work for a workflow nobody uses.
For everything that survives: export it today. Not on the twenty-ninth. Today, while the button is live and support is still answering. Then do the part almost nobody does — run the export against five real inputs from your actual work, and write down where it differs from the original. That diff is your real migration cost. The export is the easy half.
Move your prompts into your own files this week. Plain text. Versioned. If a prompt only exists inside a vendor’s console, you don’t own it, you’re borrowing it.
Build your own eval set — thirty cases from your real work, with the answers you’d accept, in a spreadsheet if that’s all you’ve got. Thirty is enough to catch a regression and small enough that you’ll actually run it. That file is the single most valuable asset in your AI stack, and it’s the one thing no vendor can retire.
And then the structural one: assume a three-year clock on every managed build surface you touch, and write the exit into the design before the pilot, not after the notice. If you can’t answer “how do we leave this in a quarter” in one sentence, you haven’t bought a tool. You’ve signed a lease with a renewal clause you haven’t read.
One more thing about this particular retirement, because it’s worth saying out loud.
The generation of tools being shut down right now was the on-ramp. It’s the reason a receptionist, a bookkeeper, a site manager, a solo founder could build something that worked. Every one of those people is now being told, in effect, that the next step requires an SDK and a CI pipeline. Some will climb. Most won’t. And the industry will call that maturity.
But a market that only lets technical people build things is not a more mature market. It’s a smaller one, wearing a nicer jacket.
The principle that has to survive is simple, and it’s older than any of this. The right to take your work with you.
A tool you can put down is a tool. A tool that keeps your work in a place you can’t carry is a landlord. And the difference between those two things isn’t in the pricing page. It’s in whether the export button produces something that runs.
Check yours this week. Then check it again in January, because the next notice is already being written.
You’ve got this. Just don’t wait for the twenty-ninth.