The Foundry workflows retirement takes effect on 1 December 2026: Microsoft Foundry’s declarative workflows, the visual designer for chaining agents and business logic, stop being supported in the portal. Your workflow definitions don’t disappear, though. Export the YAML now; Foundry keeps running it if you deploy it as a hosted agent, and Microsoft Agent Framework can load it with few changes.
What exactly retires?
Workflows were Foundry’s low-code tool: predefined sequences of steps in a visual builder, with patterns such as sequential steps, group chat and human-in-the-loop approvals, and Power Fx for expressions.
After 1 December 2026:
- The visual designer and running workflows in the portal stop being supported.
- YAML workflow definitions keep running when you deploy them as a hosted agent.
- Code-based orchestration in hosted agents is not affected.
- Prompt agents are not affected. Agents you configure with instructions, a model and tools, without code, stay.
Microsoft’s own wording: visual workflows “aren’t disappearing as a capability”. What’s going is the designer in the Foundry portal.
Step one: export the YAML
Before anything else, open each workflow in the designer, switch to the YAML view, and copy or export it. Microsoft calls the YAML “the portable artifact you carry to any of the paths below”. Until 1 December you can still edit and redeploy a .workflow.yaml under the same name, which creates a new version, but Microsoft asks you not to start new workflows this way.
Which migration route fits?
| Route | Best for | What you do |
|---|---|---|
| Microsoft Agent Framework (recommended) | Teams fine with YAML or code who want the closest match | Load the exported YAML into an Agent Framework project, often with minimal changes, and deploy it as a hosted agent |
| Azure Logic Apps | Keeping a fully visual, low-code designer | Rebuild the flow in Logic Apps, which can call Foundry agents as steps |
| Agent-to-agent (A2A) | Simple hand-offs between two agents | Connect the agents directly over A2A endpoints |
In VS Code, Microsoft’s Foundry Toolkit can also generate Python or C# from a workflow: open the workflow playground, choose Generate Code, and GitHub Copilot does the conversion. Microsoft notes this doesn’t deploy anything or guarantee identical behavior, so compare the branching, variables and approvals against your YAML before you rely on it.
What is Microsoft Agent Framework?
The successor to both Semantic Kernel and AutoGen, built by the same teams, and production-ready since version 1.0 in April 2026. It has agents, a managed agent loop, and graph-based workflows, for .NET and Python, with Go in preview. It works with Foundry, Azure OpenAI, OpenAI, Anthropic and Ollama models.
pip install agent-frameworkFor .NET, dotnet add package Microsoft.Agents.AI.
Where does Foundry stand now?
The names have moved too. Azure AI Studio became Azure AI Foundry, which became Microsoft Foundry. The old hub-based portal lives on as “Foundry (classic)”, and the Assistants API it used was sunset on 26 August 2026 in favour of the Responses API. Foundry Agent Service now has three kinds of agent:
- Prompt agents: instructions, a model and tools; Foundry runs them, no code.
- Voice agents: real-time voice, through Voice Live, in preview.
- Hosted agents: your own code and framework in a container, with Foundry’s endpoint, scaling and identity. Agent Framework, LangGraph, OpenAI’s Agents SDK and Anthropic’s agent SDK all run here.
Creating and running prompt agents and workflows has no separate charge; you pay for the model, metered tools, and for hosted agents the container compute.
We built a declarative workflow engine too
Contenox started, in April 2025, as exactly this kind of product: a declarative chain engine. Agent behavior was a JSON file, a state machine of tasks with typed handlers and branching rules, versioned in git like any other config. It ran a Telegram assistant and a GitHub pull-request bot, and it worked.
Nobody wanted it. Teams preferred to build their own workflow tech, even with a free, open-source engine available, and usually without having checked whether the automation they had in mind would pay off at all. At the same time, ready-made agents, a coding agent in the terminal or the editor, got adopted immediately, even when they were proprietary, cost money and collected data, and people kept using them through bad days and destructive commands.
So we moved on, to an agent with rules you can see instead of flows you have to design. Microsoft retiring its workflow designer, OpenAI shutting down Agent Builder and AWS closing Bedrock Agents to new customers in the same year reads like the same lesson, learned at a larger scale.
Related guides
- GenAI app builders: the 2026 picture across Google, AWS, Microsoft and OpenAI
- OpenAI Agent Builder shutdown: the OpenAI side
- Bedrock Agents Classic: the AWS side
- AI agent permissions: rules instead of flows