Microsoft announced that Azure AI Foundry’s Prompt Flow will stop receiving new features on 20 April 2026 and will be retired on 20 April 2027. It urges developers to move their workloads to the Microsoft Agent Framework. The deadline matters because the runtime images that power Prompt Flow will soon lose security updates, creating compliance headaches for any organization that keeps the service alive.

Prompt Flow was Azure AI Foundry’s visual orchestrator for large-language-model (LLM) pipelines. A “flow” is essentially a folder that contains a flow.dag.yaml file describing inputs, outputs and the sequence of nodes. The visual editor let users drag and drop components, but the underlying YAML file is what version-control systems actually track.

Why the change matters

  • Compliance risk – Without security updates, any vulnerability discovered in the runtime images stays unpatched. Regulators often label unmaintained code as non-compliant, especially when it handles credentials.

Core concepts you still need to know

Even as you plan to abandon Prompt Flow, the patterns it encouraged remain valuable. Understanding them will smooth the transition to the Agent Framework.

  • Folder-based definition – The flow’s YAML file lists every node, its type and its data contracts. Keep this file under source control; it is the single source of truth.
  • Python nodes for deterministic work – Use Python steps for parsing, validation or any logic that must produce the same result given the same input. This isolates the nondeterministic LLM calls and makes downstream testing easier.
  • Connection objects for secrets – Credentials live in a separate connection field, not in code. This design lets you swap development keys for production keys without touching the flow definition.
  • Evaluation datasets – Prompt Flow’s strongest feature was its built-in evaluation loop. A JSONL file that pairs realistic inputs with expected outputs can be reused in any framework; treat it as portable test data.
  • System-level metrics – Measuring only the output of a single node can be misleading. Combine deterministic checks (e.g., regex matches) with end-to-end metrics to catch regressions that appear only after the whole pipeline runs.

Migration roadmap

Moving from Prompt Flow to the Agent Framework is not a click-and-convert operation; it requires a rewrite of each node’s semantics.

  1. Map LLM nodes – Replace visual LLM blocks with explicit API calls wrapped in an agent class. The agent should accept the same input schema defined in the original YAML.
  2. Translate Python nodes – Shift Python code into regular functions or methods that the new orchestrator can invoke. Preserve any deterministic contracts you built around them.
  3. Rewire connections – Export the connection definitions from Prompt Flow and import them into the Agent Framework’s secret store. Verify that environment-specific values still resolve correctly.
  4. Port evaluation data – Feed the existing JSONL file into the new framework’s test harness. Adjust any metric calculations to match the Agent Framework’s reporting format.
  5. Validate end-to-end – Run the full pipeline on a sample of real traffic. Compare the final outputs against the original Prompt Flow baseline to spot drift.
  6. Iterate on orchestration – If your workflow needs resilience across process restarts, consider Azure Durable Functions or another durable orchestrator instead of rebuilding a visual flow.

What to watch next

Takeaway: Treat the April 2027 deadline as a hard stop. Pause any new Prompt Flow projects, extract the reusable pieces—YAML definitions, Python logic, connection configs and evaluation data—and start rebuilding them in the Agent Framework while security updates are still flowing. The effort now prevents a rushed, risky migration later.

Source: https://dev.to/multigrid/building-a-prompt-flow-in-azure-ai-foundry-2bbn