One framework where there used to be two
Microsoft calls it the direct successor to Semantic Kernel and AutoGen, built by the same teams. It takes AutoGen's simple agent abstractions and Semantic Kernel's session state, type safety, middleware and telemetry, then adds graph-based workflows. agentopsguide.com says AutoGen is in maintenance mode and Microsoft points new users here. If you run either one today, this is your migration path, and the docs include a guide for coming from AutoGen.
The mental model is two primitives. An agent is an LLM that takes input, calls tools and MCP servers, and produces a response. A workflow is a graph that connects agents and functions along explicit execution paths. The overview gives a rule of thumb: use an agent when the task is open-ended or conversational, use a workflow when the process has well-defined steps and you need control over order. We agree with it. Most business automation is the second kind, whatever the demo video suggests.
There's also a built-in agent for long multi-step tasks. It bundles planning and todo tracking, context compaction, file access, memory, tool approval and observability. The getting-started path on Microsoft Learn runs from a first agent through memory and workflows to that long-task agent and hosting.
Checkpointed graph workflows are the part worth switching for
The GitHub repository lists graph patterns for sequential, concurrent, handoff and group collaboration, with checkpointing, streaming, human-in-the-loop and time-travel. Checkpointing is the one that matters. A run that waits two days for an approval, or gets killed by a deploy, resumes from its last step instead of starting over.
Approvals are covered too. Microsoft's launch post says tools can be marked as needing human approval, and the framework emits a pending request you route to a UI or a queue. That's Microsoft describing its own feature, but the pattern is what we'd want for anything that writes, pays or deletes.
Observability is built in via OpenTelemetry. agentopsguide.com points out that only some frameworks ship native OpenTelemetry, and that this decides how easily you can swap tracing tools. Model support covers Foundry, Anthropic, Azure OpenAI, OpenAI and Ollama, among others, so a model change shouldn't mean a rewrite. That's what the docs say. We haven't tested it.
Where it costs more than it looks
It's single-process today. The migration guide says AutoGen had embedded and experimental distributed runtimes, while Agent Framework focuses on single-process composition now, with distributed execution planned. If you were counting on spreading agents across machines, that's a gap. Planned is not shipped.
Then there's the cloud pull. agentopsguide.com notes that Google ADK, Strands Agents and Microsoft Agent Framework each deploy most cleanly to their own cloud. The GitHub README advertises hosting on Foundry with just 2 additional lines of code, which is Microsoft's own claim. We believe it for the happy path. The sources we read state no hosting prices, so we can't give you a number, and we'd ask for one before committing.
A framework also isn't the whole stack. thefrontierwire.com argues that provider failover, spend limits and tool permissions are easier to enforce in one gateway than in ten framework configurations. We'd add that going from 1.0 to python-1.20.0 in about six months means active development, and also means you'll be reading changelogs. Pin your versions.
Against LangGraph and the OpenAI Agents SDK
thefrontierwire.com ranks LangGraph first and Microsoft Agent Framework second, and all three frameworks here are MIT licensed. LangGraph offers in-memory, SQLite and Postgres checkpointers for Python and JS, with 42.7k GitHub stars. Microsoft Agent Framework covers .NET and Python, with Go in preview, and had 13.9k stars. The article itself calls stars a weak quality signal, so don't over-read that gap.
The OpenAI Agents SDK takes a different route. It uses handoffs and agents as tools, and reaches durable execution through integrations such as Temporal, Restate, DBOS and Dapr. Microsoft's framework checkpoints natively, which agentopsguide.com also says of LangGraph. Our read: language and cloud decide this more than features do. .NET shop on Azure, pick Microsoft. TypeScript product team, look elsewhere.
Who should pick it, and who should not
Pick it if you write C# or Python, already live on Azure, and need approvals, resumable runs and traces from day one. Pick it if you're stuck on Semantic Kernel or AutoGen and want a supported way forward.
Skip it if your glue is mostly triggers and API calls. A visual tool like n8n or Make, plus a few LLM steps, is often the cheaper answer, and that's what we do in our AI workflow automation work. Skip it too if you need distributed agents across machines right now, or if your team is TypeScript-first.
Strengths, weaknesses, where it fits
Strengths
- Graph workflows with native checkpointing and resume
- Tool-approval flow for human-in-the-loop control
- Built-in OpenTelemetry tracing
- Clear migration path from Semantic Kernel and AutoGen
- Works with several model providers, not only Azure OpenAI
Weaknesses
- Single-process execution today; distributed is only planned
- Smoothest hosting path is Microsoft's own cloud
- Fewer GitHub stars than LangGraph
- Hosting prices are not stated in the sources we read
Where it fits
- Approval-gated agents on Azure and .NET
- Migrating Semantic Kernel or AutoGen code
- Long-running workflows that must survive deploys
Verdict: We'd put it in production for .NET or Azure teams that need resumable, auditable agent workflows, after a pilot and a hosting quote. For TypeScript teams or simple automations, we'd look elsewhere.
Sources
- Agent Framework documentation | Microsoft Learn · Microsoft Learn
- Best AI agent frameworks 2026: 10 open-source compared · The Frontier Wire
- 12 Best AI Agent Frameworks in 2026, Compared for Production · AgentStack
- Microsoft Agent Framework Overview · Microsoft Learn
- microsoft/agent-framework · GitHub
- AutoGen to Microsoft Agent Framework Migration Guide · Microsoft Learn
- Introducing Microsoft Agent Framework: The Open-Source Engine for Agentic AI Apps · Microsoft Foundry Blog
Want this working in your stack, not just in a demo?
We are two senior engineers who build ai workflow automation with a fixed price per written scope.