Three AI vendor shutdowns in five weeks. Every migration path is: rebuild it.

Three AI platforms that teams built real operations on are being switched off between 13 August and 14 September 2026. Flowise archived its repository. Relay.app deletes customer workspaces. OpenAI removes the Assistants API. In each case the vendor published clear guidance about what to do next, and in each case the guidance amounts to the same instruction: rebuild it somewhere else, from a description of what it did. An AI vendor shutdown doesn’t just take the tool. It asks you for a document you probably never wrote.

TL;DR: Three agent and automation platforms shut down inside five weeks in 2026, and none of them offers a working migration. Relay.app’s own docs say its JSON export is “not perfectly transferrable to other tools” and “won’t work through their native importers” — so it ships a plain-language build prompt in the export bundle instead. OpenAI says it “will not provide an automated tool for migrating Threads to Conversations.” Flowise tells users to fork the repo. The only artifact that survives a vendor shutdown is a written process definition, and the vendors are the ones saying so.

The vendor put the process description in the export box

Relay.app launched in 2021 and was positioned as a modern Zapier. It told customers on 16 July 2026 that it was winding down. Free workspaces and all associated data were permanently deleted after 15 August at 23:59 PT; paid workspaces go the same way after 14 September at 23:59 PT. TechCrunch reported on 17 August that founder and CEO Jacob Bank is joining Google as VP of Product for Chrome.

The wind-down was handled well, which is what makes it useful. Prorated refunds inside five business days. Sixty days of free access for paying customers, plus 25,000 extra steps and 10,000 AI credits a month to cover the transition. Priority support for anyone with business-critical workflows. Nobody is being mistreated here.

Now look at what’s in the export.

Relay’s documentation says the workspace export .zip contains a folder for each Workflow, Sequence, MCP Server and Table, and inside each folder: a prompt as an .md file “that you can copy and paste in other workflow builders and AI chat tools,” the run history as CSV, and a JSON file representing the exact structure of the workflow.

Then the docs explain why the .md file is there. Different workflow builders use different JSON structures, so the JSON export “is not perfectly transferrable to other tools without AI interpretation of the structure first.” You can hand the JSON to another automation tool as a prompt, “but it won’t work through their native importers.” And: “This is why we have also included a build prompt option, which we encourage you to use.”

Read that as an operator. The machine-readable file doesn’t machine-read anywhere else. The thing that actually travels between tools is the English description of what the process does. A vendor, at end of life, with every incentive to make the exit clean, concluded that the portable form of your automation is a paragraph.

That is the whole argument I’ve been making about process legibility, arriving as a file in a .zip.

The sunset calendar

Every row here comes from the vendor’s own announcement or deprecation page, checked on 21 August 2026.

Date	What switches off	The stated migration path
29 Jul 2026	Flowise: feature development ends, no further pull requests accepted	Fork the repository
13 Aug 2026	Flowise: repository archived, issues and PRs locked, npm and Docker packages deprecated	Fork the repository
15 Aug 2026	Relay.app: free workspaces and all associated data permanently deleted	Export JSON + build prompt, rebuild elsewhere
26 Aug 2026	OpenAI: Assistants API removed (announced 26 Aug 2025)	Responses API + Conversations API, migrated by hand
31 Aug 2026	Flowise: core team presence in Discord and GitHub concludes	Community forks
14 Sep 2026	Relay.app: paid workspaces, data, and stored credentials for connected apps deleted	Export JSON + build prompt, rebuild elsewhere
31 Oct 2026	OpenAI: Evals platform becomes read-only	Third-party tooling
30 Nov 2026	OpenAI: Agent Builder shuts down (announced 3 Jun 2026)	Agents SDK or ChatGPT Workspace Agents
30 Nov 2026	OpenAI: Evals platform and reusable prompts (v1/prompts) shut down	Move prompt content into application code

The Agent Builder row is the one to sit with. The visual product for building agents has a published shutdown date, announced barely a year after that category became the thing every platform had to ship.

Nobody wrote down the exception paths

Flowise is the clean case, because there is no commercial pressure in it at all.

The project was one of the most widely adopted open-source agent builders — 55,383 stars and 24,932 forks on GitHub as of 21 August 2026, under the tagline “Build AI Agents, Visually.” The repository is now archived; the last commit landed on 13 August. In the end-of-life announcement that same day, the maintainers gave their reason: “developers are increasingly relying on new coding agents to handle complex tasks,” and “the typical rigid workflow low-code approach quickly hits the limit when it comes to complexity.”

The guidance to existing users is to fork the repo and maintain it internally or through community forks. The source stays public. Nothing is being deleted.

And it still doesn’t solve anyone’s problem, because forking a 55,000-star codebase means owning dependency updates, security patches and model API churn on your own, indefinitely. That’s a standing engineering commitment taken on to preserve flows whose logic you could have written down on four pages.

This is not new for the layer. DataStax removed hosted Langflow from Astra on 9 April 2026, telling customers to use the open-source version instead. Same pattern: the canvas goes, the code is available, the running configuration is yours to reconstruct.

Here’s the part that hurts in practice. A workflow builder’s happy path is easy to rebuild — anyone can look at a screenshot and recreate five boxes in a row. What’s expensive is everything hanging off the sides. The retry that fires three times with backoff because one upstream API returns a 502 about once a week. The branch someone added in March after a customer got double-billed. The filter that silently drops records where a date field arrives as an empty string instead of null, because a vendor changed their serializer and never mentioned it. None of that is in your documentation. Some of it isn’t visible in the JSON either. It’s in the run history, and only if you know which runs to read.

Export the run history. It’s the only record of how the process behaved on the days it didn’t behave.

State was the other thing you were renting

The OpenAI Assistants API removal on 26 August is the state version of the same failure.

OpenAI gave a full year: deprecation announced 26 August 2025, shutdown 26 August 2026, replacement being the Responses API plus the Conversations API. The migration guide is direct about the limit of that generosity — “We will not provide an automated tool for migrating Threads to Conversations” — and recommends moving new threads over and handling older ones as needed, pulling messages out of each thread and converting them by hand.

A year of notice and no automated path, because Threads were shipped as a convenience and a lot of teams quietly promoted them to a system of record. I’ve written about that specific mistake in agent memory: persistence arrived as a product feature, so it never inherited the disciplines storage gets — an owner, a retention rule, a reconciliation against the system that actually owns the fact. A shutdown date is just the moment the bill for that arrives on a schedule you don’t control.

Two layers, one lesson. The logic was rented. The state was rented. Both leases end on a date the vendor picked.

What you actually own

The uncomfortable version: if the only complete description of how work gets done in your company is a diagram inside a product you don’t control, you didn’t automate that process. You leased an implementation of it, and the terms were in a footer you didn’t read.

The fix isn’t a portable format. Nobody is going to agree on one — Relay’s docs tell you exactly why, and they’re right. The fix is that the process definition lives somewhere you own, in a form a human can read, and the platform is treated as one implementation of it.

That document is short. The trigger. Every branch and what decides it. The exception paths and what happens on each. The escalation rule and its threshold. The systems each step touches and what it’s permitted to do there. Who owns the credential for each connection. Four pages for most workflows. It’s the same document you need before you can safely deploy an agent at all, and the same one that makes ownership at build time real rather than nominal. It’s also the thing a data contract is for at the seam between two systems: an agreement written down where both sides can read it, instead of a shape encoded in one vendor’s schema.

One practical note, since it catches people every time: exporting is itself a pipeline with latency. Relay’s workspace export takes up to 24 hours to build, and the emailed download link is good for 48 hours. So the actual last safe day to start is about three days before the deadline on the page, and a link that expires in someone’s inbox while they’re on PTO costs you a full cycle. Treat the export like a migration with a runbook, not a button you press on the last afternoon.

Test it the direct way. Pick your most important automated workflow and ask whether someone could rebuild it tomorrow, in a different tool, working only from what you’ve written down — without opening the platform it currently runs in. If the answer is no, the shutdown notice hasn’t arrived yet. That’s the only difference. Getting that document out of people’s heads and onto paper before a deadline forces it is the work I do.