Prompt flow in Microsoft Foundry (classic) and Azure Machine Learning will be retired on 20 April 2027. Microsoft no longer recommends it for new development. According to the documentation, the runtime containers receive no further security or package updates. The steps below are therefore aimed at teams with existing flows.
What to record first
For each flow, list the owner, business purpose, trigger, inputs, outputs, models, secrets, data sources, runtime image and deployment. Run a known test dataset and save the result. Without this baseline, it is hard to tell after a migration whether behaviour and quality have been preserved.
A migration in three steps
- 1Prioritise: Production flows with sensitive data, external actions or outdated images come first.
- 2Rebuild the workflow: Separate prompt logic, deterministic processing and tool calls. Evaluate Microsoft Agent Framework as a supported alternative rather than building a new Prompt flow dependency.
- 3Test against the baseline: Compare answer quality, cost, failure cases, permissions and logs on the same dataset. Keep a way back until cutover.
Do not start with a large rewrite before you have an inventory. A single, well-documented flow shows which dependencies and runtime differences actually matter. Align the schedule with the official retirement date.

