Skip to content
← Insights
Nov 15, 20264 min read

Google Opal is closing. Save the workflow before the link disappears

Google plans to close Opal on 17 November 2026. The useful work is the process you tested, which needs a portable definition and a replacement runtime.

A small workflow board beside an open archive box, a download arrow and a disconnected link symbol.

Google plans to turn off Opal on 17 November 2026. That changes the job for anyone who used it to build a small AI helper: save the process, then decide where it should run next. Google says existing Opal workflows will require manual migration. The stored files will remain in Google Drive, but interactive execution ends with the service. The current shutdown FAQ gives the date and the file-recovery instructions.

The appeal of the original tool was easy to understand. A team could turn a recurring task into something a colleague could open through a link. Someone supplying the input no longer needed to reconstruct a long conversation with a model. That is a useful product idea even when the service hosting the first version closes.

Describe the job independently of the editor

Opal let users connect input, generation, and output steps through a visual editor. Its documentation described outputs ranging from generated web pages to Google Drive spreadsheets. A natural-language editor could make changes to the same underlying flow. Those were concrete workflow-building capabilities, documented in Google's editor overview.

A portable description needs more detail than a screenshot of those boxes. Write down what a person supplies at the start. Specify the result they should receive. Record the instructions used at each transformation, including any reference material the result depends on. Describe the failure response beside the successful response, so the next implementation has an explicit place for missing input.

For a support-triage helper, the input might be one ticket with its relevant account context. The output could be a category plus a proposed reply. A separate condition should say what happens when the ticket contains an instruction that conflicts with company policy. The tool's owner must decide whether that case receives a draft for review or stops for a person.

This description can fit in a short document. It becomes useful when someone can implement the task without needing access to the old editor. A screenshot alone leaves too many decisions hidden inside individual steps.

Preserve examples that can reveal a bad replacement

Save a small collection of permitted sample inputs with their expected results. Include an ordinary case. Add an incomplete request that should cause the tool to ask for missing information. Keep an example where the correct result is a refusal to take an unauthorized action. Choose these cases because they expose different failure paths.

The expected answer does not have to reproduce one exact sentence. A reply draft can vary while still including the correct account limitation. The test should identify the facts that must appear, then specify any action the tool must leave to a person. That makes evaluation possible after the prompts move to a different model.

An invented claim that ticket handling falls from five minutes to one would add nothing here. A migration has to earn its own time-saving claim. Compare the total effort of completing real permitted cases, including the edits required after generation. An answer that arrives quickly can still create extra work if someone must reconstruct its reasoning or correct its assumptions.

Choose a replacement for the work it actually does

Google's September 2026 announcement presents skills in Gemini as reusable instructions for recurring chat tasks. The announcement also explains the transition away from Gems, with different schedules for personal accounts and Workspace customers. Opal workflows have their own migration requirement. A general product transition should not be read as a promise that a workflow graph will become an equivalent running skill. Google's skills announcement explains that product direction.

A task whose entire result is a draft in chat may be simple to reconstruct as saved instructions. A process that writes to another business system needs its execution path documented too. List the destination, the account that authorizes the write, and the event that tells a reviewer it finished. Keep the write disabled until the replacement passes the saved examples.

A team can also decide that the experiment has already answered its question. If nobody used the helper after its first demonstration, restoring its interface may have little purpose. Archive the definition with the reason the pilot stopped. That is a complete outcome for an experiment.

Check what was shared with the app

Google's FAQ warns that sharing an Opal also shares its Drive file, which can expose workflow details even when editor access is hidden. It also says Opal content is excluded from generative-model training, while some prompts may receive human review for troubleshooting. Those statements describe different uses of data. Neither is a reason to place credentials inside a prompt.

During the migration, inspect reference files for material the replacement should not receive. Remove obsolete account access from the old workflow where possible. Give the new runtime the permissions required for its current task. A copy of an old prompt can carry assumptions about access that no longer fit the new system.

The migration is finished when a colleague can run the permitted task, recognize a failed result, and find the person responsible for fixing it. Keep the tested definition beside that implementation. The next hosting change will then begin with a working specification instead of a dead link.

AI workflowsGoogle OpalProduct engineering