What survives the move

Overview

Why now. Somewhere in your workspace there is a saved assistant your team actually relies on — the one that holds the brief template, the house style, the checklist nobody can find in the wiki. This month one vendor put a date on it. Under its published plan for enterprise workspaces, creating new assistants of that kind ends on 26 October, the existing ones stop running on 11 December, and a migration button, targeted for late September, turns each one into a newer kind of object. The FAQ is unusually candid about what comes through the door: the instructions, and the connected apps. What stays behind: the custom integrations, the model it ran on, your old conversations, and — this is the one that catches teams — everyone’s access.

The obvious read is that this is one vendor’s product churn. The more useful read is that the vendor has written down, with dates attached, a rule that applies to every saved assistant on every platform: the prose travels, the plumbing does not. A second vendor’s voluntary copy feature carries almost exactly the same keep/lose list.

What you’ll do differently. Treat a migrated assistant as a rebuild: run your usual prompts and one hard case through it, restore whatever did not carry over, and check who can open it before anyone relies on it.

The content

Call it the prose layer. Every saved assistant is two things stacked. On top is prose: a name, a description, the instructions, the examples, the reference files. It is a text document in all but name — and increasingly it literally is one, in a published format that several products now read. Underneath is plumbing: which tools it may call, which systems it is connected to, which model runs it, and who is permitted to open it. The prose was written by you. The plumbing was granted, item by item, inside one product’s permission model, by whoever administers your workspace.

That is why the keep/lose lists look the way they do. Under OpenAI’s planned migration, the GPT’s instructions become a skill inside the new plugin, and connected apps come along as apps. Custom actions do not transfer — whoever maintains the workflow has to rebuild them. The selected model does not carry over; enterprise defaults apply. Knowledge files copy across; old conversations do not. And the replacement starts private: sharing it with people needs one permission, publishing it to the workspace directory needs another, and having had access to the original guarantees nothing.

Now Microsoft’s Copy an agent to Copilot Studio, which is voluntary. Copied: name, description, instructions, suggested prompts, icon, and the SharePoint sites and websites added as knowledge. Not copied: enterprise connectors (set them up again afterwards), embedded files (upload them again), Teams chats, meetings and emails added as knowledge, and the ability to create documents, charts and code (add it back in the agent settings). Different company, different reason for the move, same line in the same place. A grant is not a sentence. It cannot be copied into a new context, only re-issued there.

The part that fails quietly is that the prose can survive intact and the behaviour still change. Same instructions, different model, different harness around it. OpenAI’s own guidance: “A migrated plugin may respond differently. Compare familiar prompts and at least one harder case.” The familiar prompts tell you the words came through. The hard case tells you whether the plumbing did — because the hard case is the one the old assistant only got right thanks to an integration or a reference file it no longer has.

Two more things go wrong silently. After migration the original keeps running but becomes read-only, and its creator cannot delete it — so for a while there are two of them, one frozen, and a colleague with the old bookmark may be using the stale one. And migration carries over none of the old sharing settings, so the people who used to open the assistant are exactly the ones who will find the replacement closed.

Microsoft’s page gives the same instruction in plainer words: “Test your agent in Copilot Studio before you publish it. After you publish, decide whether to delete the original agent and communicate migration steps to users you shared the agent with.” Rebuild, test, restore, tell people. Not migrate and hope.

Try it

Do this whether or not a banner has appeared. Pick one saved assistant your team depends on — a custom GPT, an agent, a Gem, a Project with standing instructions — and write its rebuild sheet. Five lines:

1. Instructions — copy the text out into a document you own. This layer travels; keep your own copy anyway.
2. Reference material — list every file it was given.
3. Integrations — name each action, connector or app it can reach. These are the ones that will need to be re-granted.
4. Model — which one, and whether you chose it deliberately.
5. Access — who can open it today, and who must be able to open the replacement.

Then run three of your usual prompts through it, plus one hard case — the request that only works because of a file or an integration — and save the outputs. If migration or copy is already available in your workspace, run it and compare the replacement against that baseline line by line. If it is not available yet, you now have the baseline waiting for the day it is.

Where availability varies: migration and sharing sit behind workspace permissions you may not hold. Check what your workspace has enabled — and if sharing the replacement is blocked, the ask to your admin is specific: the share or publish permission for the new object, for the named people on line five.

Additional reading

• Custom GPT retirement and migration FAQ — OpenAI Help Center, article 20001519. The dates, the keep/lose list, and the test checklist, in the vendor’s words. The page refuses automated fetching; read it in a browser.
• ChatGPT Enterprise and Edu release notes, 11 September 2026 — the admin-notice entry; “Dates may change.”
• Copy an agent to Copilot Studio — Microsoft Learn, updated 29 July 2026. The voluntary version of the same move, with its own copied/not-copied table.
• Agent Skills specification — agentskills.io. The published format the prose layer is increasingly saved in: a folder, a SKILL.md, a name and a description.

Editor’s note

Plenty of people have never had to deal with this, because someone else builds the assistants they use. If that’s you, expect it to change soon, at least for the assistants you build for yourself. A move like this can change the output you expected more than you’d think, so test before you treat the job as done.