Workflows, and where Microsoft Agent Framework fits

Lesson 4 of 5 in Microsoft Foundry Agent Service, Three Ways to Run an Agent.

Foundry workflows are the declarative multi-agent story: a visual builder plus YAML configuration, with templates for human-in-the-loop, sequential and group-chat patterns, branching logic and variable handling — no code required. They have been preview throughout.

They are also being retired. This is the single most important date in this module, and the reason to read a course before a demo.

Where multi-agent orchestration goes after the designer
OptionWhat it isWhen it is the right answerWhat you give up

Microsoft Agent Framework

Open-source SDK and runtime with built-in orchestration patterns, graph-based workflows, and declarative YAML.

Anything with real branching, checkpoints, or logic you expect to test and version like code. The primary documented path.

A designer your non-engineers could edit. Orchestration becomes an artifact in your repository — which is the point.

Azure Logic Apps

The existing enterprise integration and workflow engine, with its connector estate.

Process automation where the agent is one participant among many systems — approvals, tickets, ERP hops, scheduled triggers.

Agent-native abstractions. You are orchestrating steps, and the agent call is one of them.

Direct A2A connections

Agents calling agents over the A2A protocol, using Foundry’s incoming A2A endpoint (preview).

Lightweight hand-offs between agents that each own their own logic — no central coordinator worth building.

Central visibility of the whole plan. Nobody holds the flowchart; you get emergent choreography and the debugging bill that comes with it.

So what is Microsoft Agent Framework, and why does it keep appearing on both sides of the platform boundary?

It is the direct successor to both Semantic Kernel and AutoGen, built by the same teams — AutoGen’s simple single- and multi-agent abstractions plus Semantic Kernel’s enterprise machinery (session-based state, type safety, filters and middleware, telemetry), with graph-based workflows added. Microsoft’s own phrasing: the next generation of both. Both predecessors remain supported, but investment has moved. Announced on 1 October 2025 as a fully open-source SDK and runtime, with open standards — MCP, A2A, OpenAPI — as one of its four stated pillars.

Its relationship to Agent Service is the part people get wrong. Agent Framework is not the service, and the service does not require it. It is a library you can run anywhere; Foundry is a place to run agents. They meet in exactly two places: Agent Framework is one of the supported ways to build a hosted agent, and it has a Foundry integration for driving prompt and hosted agents from framework code.

The orchestration patterns you get for free

Agent Framework ships built-in multi-agent orchestration patterns: Sequential, Concurrent, Handoff, Group Chat, and Magentic — a manager agent that dynamically coordinates specialists. It also offers “workflows as agents”, exposing a whole workflow through the standard agent interface.

That last one is the composition trick worth remembering: if a workflow is an agent, a supervisor cannot tell whether the thing it delegated to is one model or a graph of twelve. Same interface, replaceable implementation.

The four areas of the framework

Microsoft organizes it as: Agents (model clients for Microsoft Foundry, Anthropic, Azure OpenAI, OpenAI, Ollama and more), Harness Agent (an opinionated agent for long multi-step tasks with planning and to-do tracking, context compaction, file access and memory), Workflows (functional and graph-based), and Integrations.

Foundational building blocks: model clients, agent sessions for state, context providers for memory, middleware, and MCP clients. Human-in-the-loop pauses, checkpoints with resume, and OpenTelemetry-based observability are built in rather than bolted on.

Language support — check before you commit

Microsoft Learn documents Python, C# and Go pivots. As of this pack’s September 2026 check, the Go surface was public preview with some features absent (declarative agents, RAG, CodeAct, functional workflows), and the FoundryAgent integration for prompt and hosted agents was not yet available for Go. Python ships as agent-framework-foundry; the .NET package Microsoft.Agents.AI.Foundry was prerelease at that check.

Treat “is Agent Framework GA?” as a per-language question and answer it from the current docs, not from a blog post — including this one.

Getting a multi-agent design onto a surface that will still exist next year

  1. Multi-agent orchestration needed

    Two or more agents, or one agent plus real branching and hand-offs.

  2. Already built in the visual designer?

    Retirement date: 1 December 2026. The designer and in-portal execution end then.

  3. Export the workflow YAML

    Exported YAML remains runnable — the asset survives, the drag-and-drop surface does not.

  4. Deploy the YAML as a hosted agent

    Workflows are no longer a separate agent type; a workflow now runs as an agent like anything else.

  5. Needs code-level control, tests, checkpoints?

    Branching you must unit-test, resumable long tasks, custom hand-off logic.

  6. Agent Framework, run as a hosted agent

    Sequential, Concurrent, Handoff, Group Chat, Magentic — plus workflows-as-agents, checkpoints and OpenTelemetry.

  7. Logic Apps, or direct A2A hand-offs

    Logic Apps when the agent is one participant in an enterprise process; A2A when peers just need to call each other.

  8. Versioned, observable, supported

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.