Microsoft Agent Framework Is The Enterprise Agent Stack Builders Need To Watch
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
Microsoft Agent Framework matters because it is not another thin chat wrapper. It is Microsoft pushing agent apps toward something builders can actually operate: agents, workflows, checkpoints, human approval, observability, MCP, and deployment in Python and .NET.
If you are building a real business process with AI, this is the part to pay attention to. The future is not one model answering one prompt. It is structured work that can pause, resume, route, get reviewed, and recover when something breaks.
Quick takeaways on Microsoft Agent Framework
- Best fit: builders making agent apps, multi-agent workflows, and enterprise-style automation in Python or .NET.
- Why it matters: Microsoft Learn says Agent Framework supports agents and workflows with state management, context providers, middleware, MCP clients, checkpointing, and human-in-the-loop support.
- Fresh signal: the Python 1.7.0 release shipped on 2026-05-28 with HarnessAgent, A2AAgentSession, Foundry deployment APIs, Foundry Toolbox MCP invocation support, and DevUI streaming memory fixes.
- GitHub signal: the microsoft/agent-framework repo showed 10,821 stars, 1,796 forks, 911 open issues, and a push at 2026-05-28T14:15:44Z when checked on 2026-05-28.
- Real talk: this does not replace OpenClaw. It solves a different layer.
What is Microsoft Agent Framework?
Microsoft Agent Framework is a framework for building, orchestrating, and deploying AI agents and multi-agent workflows with support for Python and .NET. Microsoft describes it as a way to build both agents for open-ended tasks and workflows for well-defined multi-step processes.
That distinction is the whole game. A lot of AI builders treat every problem like it needs a smarter chatbot. It usually doesnt. Some parts should be agents. Some parts should be deterministic code. Some parts should be a workflow with a checkpoint and a human approval gate before anything touches a customer.
The Microsoft docs list the pieces that matter for production agent work: state management, context providers, middleware, MCP clients, checkpointing, and human-in-the-loop support. Those are not flashy demo words. Those are the boring rails that keep an agent from turning into a random button with a model behind it.
Why Microsoft Agent Framework matters for real agent apps
The important part is workflow control. Microsoft Agent Framework gives builders a path to move from one-off agent demos into restartable, observable business processes.
Think about a support ticket workflow. The agent can read the ticket, classify it, draft a GitHub issue, ask for human approval, create the issue, and draft a customer reply. The weak version is one prompt trying to do everything at once. The stronger version is separate steps with clear inputs, outputs, checkpoints, and review gates.
That is where Agent Framework starts to make sense. Microsofts workflow samples include basic pipelines, parallel fan-out and fan-in, checkpointing, human-in-the-loop review, and resume after interruption. That is closer to building a reliable business process than chatting with a model and hoping it remembers what happened five minutes ago.
| Layer | What it does | Why builders should care |
|---|---|---|
| Agents | Handle open-ended reasoning and language tasks. | Good for review, planning, extraction, drafting, and judgment-heavy work. |
| Workflows | Control known multi-step processes. | Good for support triage, approvals, issue creation, reporting, and repeatable ops. |
| Checkpoints | Save progress so work can resume after interruption. | Critical when a workflow touches customers, databases, files, or paid APIs. |
| Human-in-the-loop | Adds review gates before risky steps. | Lets the agent move fast without pretending every action should be automatic. |
| MCP clients | Connect agents to external tools and context. | Turns the agent from a text box into something that can operate inside a real stack. |
What changed in the Python 1.7.0 release?
The Python 1.7.0 release is a strong signal that Microsoft is pushing this framework toward background agents, A2A sessions, MCP invocation, hosted deployment, and checkpoint reliability. The release was published on 2026-05-28 at 10:46:31 UTC.
The added items are the interesting part. The release notes list HarnessAgent and a background-agents harness provider, A2AAgentSession with referenced task IDs and input-required support, experimental prompt-agent conversion and deployment APIs, Foundry Toolbox MCP invocation support, and hosting samples.
The fixes matter too. The release notes mention a DevUI streaming memory growth regression fix, hosted checkpoint fixes, OpenAI streaming metadata fixes, and other reliability work. That tells you where the pain is. Agent systems are not hard because the model cant write a paragraph. They are hard because streaming, memory, checkpoints, hosted state, and review gates all have to behave when the workflow is no longer a toy.
How to think about Microsoft Agent Framework vs OpenClaw
Microsoft Agent Framework and OpenClaw should not be treated like fake rivals. They sit at different layers of the agent stack.
OpenClaw is better as a personal and business operator layer. It can live across channels, memory, local machine tasks, Telegram, cron, daily workflows, and the messy places where a founder actually works. Microsoft Agent Framework is better when you are building a structured agent app with Python or .NET and you need orchestration patterns, checkpoints, observability, and deployment rails.
The clean mental model is simple. Use OpenClaw to run the operation. Use Microsoft Agent Framework to build a specific agent application inside that operation.
| Use case | Better fit | Main tradeoff |
|---|---|---|
| Personal operator across Telegram, local files, memory, and recurring tasks | OpenClaw | Less focused on being a Python/.NET app framework. |
| Enterprise agent app with typed workflows and deployment patterns | Microsoft Agent Framework | More app-builder oriented, less like a daily operator. |
| Support ticket to GitHub issue workflow with approval gates | Both together | OpenClaw can supervise the operation while Agent Framework handles the app workflow. |
| Teaching non-technical founders how agents fit into a business | OpenClaw first | Starting with framework code too early can distract from the actual workflow design. |
When Microsoft Agent Framework is the wrong tool
Microsoft Agent Framework is not the first thing I would hand to a founder who just wants an agent to help run the day. If the job is reminders, Telegram replies, memory, daily content checks, local automations, and lightweight business operations, starting with a framework can be overkill.
It is also the wrong place to hide unclear thinking. If you dont know the workflow yet, adding graph orchestration does not magically fix it. You still need to define the steps, the handoffs, the failure modes, and the human review gates. Otherwise you just built a more complicated way to be confused.
The best use case is when the process is already real. Support triage. Research pipelines. Content review. Data enrichment. Sales follow-up. Internal operations. Anything where the business process exists and the agent needs rails.
How builders should use it without getting lost
Start with one workflow, not the whole company. Pick a process with a clear input, a clear output, and one risky step that needs review.
For example, support-to-GitHub is a good test. The input is a customer support ticket. The output is a GitHub issue plus a customer reply. The risky step is creating the issue or replying to the customer. That is where the human-in-the-loop gate belongs.
Then split the system into boring parts:
- Classify: What kind of ticket is this?
- Extract: What product, error, customer, and reproduction steps matter?
- Draft: What GitHub issue should be created?
- Review: What should a human approve or edit?
- Execute: What tool call creates the issue?
- Reply: What should the customer hear back?
That is how you avoid the trap. Dont build an agent because agents are cool. Build a workflow because the business has repeatable work that deserves a better machine.
FAQ: Microsoft Agent Framework
What is Microsoft Agent Framework?
Microsoft Agent Framework is a Microsoft framework for building, orchestrating, and deploying AI agents and multi-agent workflows in Python and .NET. It combines agent abstractions, workflow control, state management, context providers, middleware, MCP clients, checkpointing, and human review support.
What is the difference between AutoGen and Microsoft Agent Framework?
AutoGen focused heavily on multi-agent conversation patterns. Microsoft Agent Framework is the broader production direction that pulls together AutoGen-style agent work with Semantic Kernel-style enterprise features, structured workflows, deployment patterns, checkpoints, and observability.
Should I use Microsoft Agent Framework or OpenClaw?
Use Microsoft Agent Framework when you are building a structured agent app in Python or .NET. Use OpenClaw when you want an operator across channels, memory, local tasks, Telegram, cron, and daily business workflows. The best answer for serious builders may be both.
Does Microsoft Agent Framework support human approval?
Yes. Microsoft Learn lists human-in-the-loop support, and the public workflow samples include HITL review examples. That matters because real agent systems need review gates before they touch customers, money, databases, or production tools.
The bottom line
Microsoft Agent Framework is one of the clearest signs that agent apps are moving from demos into structured, observable, restartable workflows. That is the shift builders should watch.
The winning stack will not be one magic agent. It will be the right operator layer, the right workflow framework, the right memory, the right approval gates, and the discipline to make agents do real work instead of just talk.
If you want help figuring out which agent stack actually fits your business, join Shipping Skool. We build the workflows, test the tools, and turn this stuff into real systems.
Chapters
Transcript
Read the full transcript
0:00So, Anthropic just bought the plumbing that makes AI agents useful, and in this video, I'm going to break down exactly what Stainless is and what it means for builders and AI agents going forward. So, this is the plumbing, the infrastructure, the architecture that makes AI agents go, and I think this is one of those stories that looks really boring for a few seconds, and then you realize it explains where this whole AI agent race is going and ending up. So, in this video, I'm going to be talking about what Anthropic did and how they acquired Stainless and why SDKs and MCP servers matter way more than people think and what builders should do with this information. So, the official Anthropic announcement from May 18th,
0:47just last week, and Anthropic says it's acquiring Stainless. It's basically a company that generates SDKs, CLIs, and MCP servers from API specs. Anthropic says Stainless has, you know, powered every official Anthropic SDK since the earliest days of Anthropic API. And I'm talking every single Anthropic SDK. You know, it says here, "The frontier of AI is shifting from models that answer to agents that act." And agents are only as capable as the systems they can reach. So, last week, they acquired Stainless. It was founded in 2022, and Stainless has literally powered the generation of every single official Anthropic SDK
1:32since the early days, which is pretty impressive. So, we're talking TypeScript, Python, Go, Java, more languages, the libraries developers import when they want to connect their apps to Claude, and command line tools. That's pretty simple, right? The connectors and the stuff that actually turns an API from a webpage into something a real builder can actually use inside of a product. I love that. And the part that most people are kind of miss, but pay attention here, is Anthropic did not say agents are only as smart as the model. They said agents are only as capable as the systems they can reach. So, you need to check that. I have the source link down below. Just overview it. There's a huge shift going on because for the last year, most people have basically been arguing about
2:19model benchmarks. I mean, benchmarks are great, but have you ever seen a benchmark of a new LLM come out? You go and download the LLM or you import it into your IDE, and then it just sucks. It sucks at code review. It sucks at refactoring your code base. It sucks at tool calling. And you're like, this scored a 97 on SweBench. What is going on? That's why benchmarks are amazing until you actually use it. We need real-world production use for these models. And everyone's been arguing about benchmarks. They get the clicks, but that's not the real point here for builders. It's always been Claude versus GPT or Codex versus Claude Code. And lately it's been Hermes versus Open Claude. It's like, who cares? People have even been saying local models are taking over. It really doesn't matter. It depends on what you're building, right? And yes, the model matters here,
3:05but just not as much as it people think, right? If an agent cannot reach your calendar or your database or your CRM or your payment processor, more importantly your repo, you know, it is still basically just a smart intern trapped in a glass box. And that's why Stainless acquisition matters here. Anthropic says Stainless turns an API spec into SDKs across all languages: TypeScript, Python, Go, Java, and more. And it also says hundreds of companies rely on Stainless to generate SDKs, CLIs, and MCP servers. So, that means Stainless is not just developer convenience. It's distribution for builders. And it's adoption. It is the bridge between a company saying we have an API and a
3:51builder actually shipping something with an API. So, Anthropic owns that bridge now and TechCrunch reported a sharper version of the story. So, basically TechCrunch came out and said Stainless was a used by OpenAI, Google, Cloudflare, and then Anthropic told TechCrunch it will wind down all hosted Stainless products including the SDK generator. So, this is not just about Anthropic buying a helpful internal tool. This is Anthropic pulling agent infrastructure closer within Claude and that's why I wanted to talk about this story because it's relevant to us builders. And if you're building with AI agents, you really need to pay attention to this because the next wave is not the one chatbot, it's agents that can actually safely run real systems. So, Anthropic also said it created MCP to
4:37make agent connectivity more possible. So, MCP basically stands for model context protocol. It is the open protocol for Anthropic and they've been using it to push agents so they can connect to tools and data sources in a cleaner way. That's the real play. And a model by itself this is just a brain and MCP is kind of like the hands. It's the way it works with these tools and skills. SDKs give builders an easy way to put those hands into real apps. I hope this is making sense. CLIs, command line interfaces, let operators use those hands from a terminal. I know it's a lot of jargon but these are really important concepts you need to understand if you want to build with AI. And if you combine all those together, oh my, you
5:22get a beautiful piece of the puzzle coming together. You get agents that do not just chat with you but they actually work and do work autonomously for for you, agentically, right? And they can read records and they can like create tickets and update documentation. They can call APIs, they move data, they can run repeatable workflows, and that is what people are actually paying for in businesses, not just another wrapper. We're past that point. And not just another pretty chat UI chatbot, we're over that. We want actual work getting done reliably. And here's the builder angle here if you're building with AI, which I'm sure you are. So, if you're building with AI or you're building an AI tool right now, your moat is probably not going to be your prompt. And I know
6:07people don't want to hear that. I don't want to hear that. But prompts are getting copied instantly. Very easy to go into Claude code or Google AI Studio or Chat GPT and come up with a really good-looking prompt. Even workflows are getting copied right now. And UI patterns are getting copied. I mean, every vibe coded website looks the same. It's annoying. Even model access is getting more equal because everybody is routing through their same few frontier providers. The part that gets harder to copy is the integration layer. The exact way your agent connects to messy business systems, the exact permissions model, the exact approval flow, and the exact logging. So, that's why I wanted to talk about this. This is a really important back-end infrastructure on what Anthropic is doing. And that is
6:53where the money is going, right? Anthropic is building and buying Stainless, and that's a huge signal here. They're not just trying to make Claude code smarter or the whole Claude ecosystem smarter, they are trying to make Claude easier to attach to basically the rest of the software world. So, if you read this line in particular from Anthropic again, they say in this article here published on the 18th that agents are only as useful as what they can connect to. And I 100% believe that. That should be written above every AI builder desk right now. Get it tattooed on your forehead. I don't care. This is important because most people are still building like the model is the product. The model is not the product. The finished job, the workflow, the agentic workflow, the cron
7:39job that comes on every day on a repeatable task, That is the finished product. That's what people are buying. If your agent helps a reseller list 200 items faster, that is the product. I built Easy Flip. It is a app that I built last year in Cloud Code. It's generated over $6,000 in revenue, and I was just solving a small problem. How can I help people list items on eBay faster with the help of AI? That's the product. For another example, another app that I made for AI agents is a real estate AI photo editor. So, if your agent that was created helps real estate agents edit photos and publish listings to the MLS faster, that's the product. So, if your agent helps a small business owner answer leads or send follow-ups or update the CRM or it can book calls,
8:25that's another product. The model is just the engine. The workflow is the car, and the integration layer is the road. No road, no trip. That's sad. But, what should builders actually do with this information? I mean, come on. I've been blabbing everywhere. What should they actually do about this? I think the first thing you need to do is stop thinking of APIs as developer docs only. Your API is now an agent service. If an human developer can use your API, an agent should be able to use it, too. If you're building a product and you don't give API access, what are you doing? Build your product for agents, not for humans. That's the next revolution here. And that means having clean documentation and stable examples, simple authentication, and good error messages. You need predictable objects,
9:11small scoped permissions, and source for the MCP and SDK. And if you're building a SaaS right now, you need to start asking one question to yourself. What would it take for an AI agent to use my product without human babysitting? Hmm. Not recklessly, either, and not with full account access, but just safely. I mean, Stripe just did that earlier this year. I mean, it was like a month ago, where they actually made money for AI agents, right? And that's going to be huge. So, I'm asking, can your agent read what it needs to read? Can it write only what is allowed to write? Can it ask for approval before anything risky? And can it leave an audit trail? Can it recover something if it fails? That is where the agent economy essentially is and where it's going right now. The second thing you need to ask yourself is
9:56you need to learn how MCP works like right now. Not because MCP is magic and it's not, but because it is becoming one of the most default ways people think about tool access for agents. Entropic says it created MCP to make the agent connectivity possible and the stainless acquisition is directly tied to SDKs and MCP server tooling. So, if you're inside the community right now, inside of our shipping school community, this is the kind of thing I wouldn't ignore. You do not need to become some protocol expert, but you should understand the basics like what is a tool, what is a resource, and what does the server expose, and what should never be exposed. Those are the things you need to know. Like, how do approvals work and how do you keep
10:41secrets out of model contexts? I mean, that's exactly why we've kind of redid the entire shipping school community is we have this course outline being built. It's a 24-week sprint of answering all those questions. What is an API? What is an AI agent? How does machine learning work? How does AI inference work? What is machine learning? How do AI agents communicate with each other? What is MCP? What is authentication? We go over what is Supabase? What is Neon? What is PostgreSQL? What is vector embeddings? All of the technical details to help you become the builder you need to be in the AI era. I'll put the link to the community down below. These are going to be courses that are released four times a week and you're seeing the syllabus right before your eyes. It's a 24-week sprint of every single thing you need to
11:26know about AI to hit the ground running to build with AI agents. I really think once AI agents can connect to everything, the winners are going to be the people who know how to connect to other tools safely. That is it. I believe the third thing we really need to be thinking about is if you have a boring business, good. That means no one else is going to do it and you're going to have less competition. I know that sounds weird, but stick with me. Boring businesses are full of weird workflows. Things like invoices. God, I'm already bored. Listings, customer messages, quote requests, scheduling inventory, even follow-ups is pretty boring. But every one of those has tools, APIs, spreadsheets, portals, and old processes around it that you could build a business on. That is agent fuel. The big
12:12companies are fighting over that infrastructure, but the small builders can win by turning that infrastructure into specific workflows. Not general AI, but specific outcomes. An agent that reconciles contractor invoices, that is huge money. An agent that turns product photos into marketplace listings, money. An agent that watches failed builds and opens a fixed PR, oh you're swimming. An agent that checks lead emails or drafts a reply or updates a pipeline, that's the game we're playing here. And if product and if your product basically has already an API, you need to make it agent readable. And if your product doesn't have an API, I need to know why. Like come on. Because we are moving into a world where the buyer might be a
12:57human, but the daily user is probably going to be an AI agent. This isn't hype anymore. This is actually happening. That is what these acquisitions are pointing at. Anthropic says hundreds of companies rely on Stainless for SDKs, CLIs, and MCP servers. I mean, I have the source down below. Read it. TechCrunch says the acquisition takes a key infrastructure supplier out of the hands of Anthropic competitors and says Anthropic will wind down hosted Stainless products. Put those together and Anthropic wants tighter control over how Claude connects to the world. That does not mean that every builder should panic, but it does mean we should be honest about the direction of where this is going. AI companies are not just competing on chat anymore. They're competing on distribution. They are competing on developer experience and
13:43they're competing on tool access. Essentially, they're competing on who becomes the default way agents touch software. And that is why Open Claude matters here, too. I know I need to put this in here somewhere. Open Claude is not just another chatbot. The Open Claude readme describes it as a personal AI assistant you run your own devices with channels like Telegram, Slack, and Discord. And that means Open Claude's already built around the idea that AI agents live where the work happens, like messages and calls and files and tools and skills and cron jobs. That's what makes it so appealing. Getting it up and running is another thing. So, the lesson from Anthropic and the Stainless acquisition is not that everyone should only use Claude. It's the lesson is that agent builders need to own their own workflow layer. So, if you're using Open
14:28Claude, Hermes, Codex, Claude code, or whatever, I don't care. Just start thinking less like a prompt engineer and more like an operator. Like, what systems does my AI agent need to work? And what permissions does it need? What should it never touch? What does it need to ask approval for? And what does it leave proof like an audit trail? And where does it fail safely so you know that it fails? That is the difference between a demo and an actual business. And that is why I keep saying that money is not in generic AI. The money is in the workflow automation and the money is in the boring integrations. Basically, the money is in taking a painful job that happens every day and running it with supervision. Anthropic buying Stainless is a loud signal that the big labs know this, too. And they know the next battle is not just bigger context
15:14windows or better benchmark screenshots, even though those get all the clicks. Like, I mean, look on Twitter. It's like every Every a benchmark release, it's like 150,000 uh views and 3,000 likes and the LLM sucks. Can't even tool call properly. What's going on? The money's in the connection. It's in the action. It's in the reliability. It is in agents reaching the tools where the real work lives. So, if you're builder, here is the move. You need to pick a business workflow and map every tool it touches. Find the API docs. Find the MCP server if one exists. If one does not exist, build it. Then building the smallest agent that can compete and complete one real job with an approval and logs, that's money. Just have it do one job. A job someone would pay to stop doing
16:00manually. That's money. That is how you turn this news into actually something useful. Focus on one thing that's simple. Automate it. Maybe it's something that's already in your business that you hate doing. Bu- build it for yourself and then sell it. Those are the road signs and I can't drive the car for you. I can't build for you. You have the tools. You need to actually take action. Start messy. Start before you're ready. Even if you don't know how to use these tools, how do you think I learned how to use them? By just building and breaking things. So, that's the road sign and it's very clear. The AI agent race is moving from intelligence to access. And the best agent is not just the one that thinks the best, it's the one that can safely reach the right tool at the right time with the right permission and finish the actual job. If you want help building these actual workflows, like I said, I
16:45showed the syllabus earlier, join us in inside the Shipping Community. Shipping School Community. We have over 150 members. And like I said, four live calls a week and we are building up the course library so you can go back. If you can't make those calls, you can know what an MCP is, how an API works, what is AI. You can even learn how to learn, learn how to install local LLMs if you are a local LLM enthusiast and you have the hardware to run it. We teach technical deep dives as well. So, like we're not just sitting around and debating all day. We're actually learning and building things and we're here to help you build them. So, we're building agents, we're building automations, tools, real stuff that real businesses actually need. I'll put the link down below. You need to check it out. If you haven't subscribed to the video or the channel already, what are
17:31you waiting for? We'll see you in the next one. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Microsoft Agent 365 Shadow AI: Why OpenClaw Builders Need Governance Now
Microsoft Agent 365 now detects Shadow AI activity like OpenClaw. That is validation, but it also means agent builders need permissions, logs, approvals, and governance now.
Paperclip AI Agent Orchestration: Run Your Agents Like a Company
Paperclip is an open-source, self-hosted Node.js platform that runs your AI agents as a company, with org charts, per-agent budgets, goal tracking, governance, and a dashboard. It works with OpenClaw, Claude Code, Codex, and Cursor, and you install it with npx paperclipai onboard --yes.
OpenAI Symphony: Why Coding Agents Need Tickets, Not Bigger Chat Windows
OpenAI Symphony shows the next AI coding bottleneck is not model intelligence. It is orchestration, task clarity, review, and human attention.
