Move your AI agents from prototype to production using Microsoft Foundry Agent Service.
Deploy directly from your local environment, run with secure identity and scoped permissions, and monitor every interaction so you can debug, improve, and scale without losing control. Publish agents into the tools your team already uses and ensure every action is traceable, governed, and isolated.
Ground your agents in real work and business data to generate outputs that are actually useful. Pull from emails, meetings, and operational systems to create personalized insights, documents, and presentations. Build faster with familiar tools and frameworks, then manage performance, cost, and quality across all your agents as they scale.
Jeff Hollan, Partner Director, AI Agent Services, shares how to operationalize AI agents across your organization — from deployment to real-world impact.
Control what your agent can access.
Assign scoped permissions and identities so every action is traceable and compliant. See how it works in Microsoft Foundry.
Scale agents without losing visibility.
Monitor performance, conversations, and health in one place with Microsoft Foundry. Check it out.
Pull insights from across systems.
Prepare faster and make better decisions. Act with full context, not guesswork using Work IQ, Foundry IQ, and Fabric IQ.
QUICK LINKS:
00:00 — Build single and multi-agentic workloads
00:44 — Build agents at scale with Foundry
01:33 — Demo: Sales meeting preparation agent
03:32 — How it works
04:48 — Access controls
05:44 — Publish the agent
06:23 — Direct integration with Microsoft 365
07:26 — Work IQ, Foundry IQ, & Fabric IQ
10:24 — Agent creation
11:21 — See what’s happening in the code
12:54 — Manage performance
13:56 — Wrap up
Link References
Go to the Microsoft Foundry to build your first project at https://ai.azure.com
Check out https://github.com/microsoft-foundry
Unfamiliar with Microsoft Mechanics?
As Microsoft’s official video series for IT, you can watch and share valuable content and demos of current and upcoming tech from the people who build it at Microsoft.
- Subscribe to our YouTube: https://www.youtube.com/c/MicrosoftMechanicsSeries
- Talk with other IT Pros, join us on the Microsoft Tech Community: https://techcommunity.microsoft.com/t5/microsoft-mechanics-blog/bg-p/MicrosoftMechanicsBlog
- Watch or listen from anywhere, subscribe to our podcast: https://microsoftmechanics.libsyn.com/podcast
Keep getting this insider knowledge, join us on social:
- Follow us on Twitter: https://twitter.com/MSFTMechanics
- Share knowledge on LinkedIn: https://www.linkedin.com/company/microsoft-mechanics/
- Enjoy us on Instagram: https://www.instagram.com/msftmechanics/
- Loosen up with us on TikTok: https://www.tiktok.com/@msftmechanics
Video Transcript:
- AI agents are gaining traction everywhere right now, but moving from experimentation to production, especially in enterprise environments is where most people get stuck. So to solve for this, today we’ll get hands-on with the Microsoft Foundry Agent Service, a platform which lets you bring in your own agents using your preferred tools and host them with built-in enterprise controls, measurability, and discoverability, and the powerful open-source Microsoft Agent Framework that’s uniquely designed to make it easier to build both single and multi-agentic workloads with orchestration. And joining me to demonstrate all this is resident developer expert, Jeff Hollan. No stranger to Mechanics. Welcome back.
- I’m so excited to be back.
- Yeah, so it’s been a while. It’s good to have you back on. So these two services that we’re covering today, both are for hosting and building agents themselves. So what’s driving all this?
- What’s driving this is something that we are hearing constantly, which is even though it’s gotten easier to build agents, it’s hard to deploy them safely and reliably across the enterprise, especially considering that a lot of what we see getting built has moved past the small pilot phase. Some agents might be chat experiences used by thousands of employees while others run behind the scenes, sometimes integrating with mission-critical systems. These all need foundational capabilities, like identity and access controls, private connectivity, along with agent and fleet-level telemetry and tracing, which is complex to stitch together by yourself. And so our Foundry services help you build agents that will run securely and at scale with full visibility.
- So I’d love to see an example of this. Did you come prepared?
- Of course, I came prepared. Let’s jump into it. So what I’ve helped walk through here and built is a sales meeting preparation agent. This is the kind of thing that a sales team would use to get ready for customer meetings. Now I already have my code written here and ready to go. I’ve used my framework of choice. In this case, this is the Microsoft Agent Framework written in Python, but you could bring your own framework and language. And you can see that I’ve defined quite a few tools here, some middleware logic and even a workflow. Now, all of these details we’re going to jump into in a bit, but importantly for now, I just want you to know that this is all running locally. It’s ready to go. I’ve built it out. Now the question often comes, how do I take something like this but get it deployed? How do I make sure that it can run in a secure and scalable way that’s compliant and safe across my entire enterprise? And Foundry makes this incredibly easy. So right here in Visual Studio Code, if I expand out the AI Toolkit extension, I can simply hit deploy to hosted agents. This gesture takes my agent as I’ve written it, packages it up, and deploys it inside of Foundry as a hosted agent. So why would I want this inside of Foundry? And I want to walk you through some of what lights up the moment that I do that. So here in the Foundry portal, you can see that this is the same agent that I was just looking at locally, but now it’s running inside of Microsoft Foundry. So let’s go ahead and call this agent from the playground so I can show you all the type of capabilities that it has in action. So I’ll ask it, what important meetings do I have this week? Now I’m actually using some of the more modern agent patterns here. So my agent is actually executing inside a secure sandbox or microVM. So you can see this agent is actually starting to think and work through the problem, looking at my calendar. It has the ability to write and execute code, very much like powerful coding agents like GitHub Copilot CLI or Claude Code. Now, while it runs, I’ll describe a little bit of how this works behind the scenes. First, as soon as you deploy the agent, it gets its own unique ID assigned from Microsoft Entra. The ID makes it so that any action the agent makes, like looking at my calendar, is traceable back to the agent. And it allows the agent to autonomously access resources directly with its own scoped permissions, or the agent can act on behalf of a human in the loop using the user’s permissions instead. And on top of all of that, for each user that invokes an agent session, Foundry automatically spins up a secure microVM, which is an isolated sandbox. So now if I ask a question and another salesperson asks another question at the same time, because we each have our own agent instances, the information from each of our sessions can be read, written, and stored in its own dedicated space. Additionally, for every interaction, the service looks at any policies or guardrails set by your organization. This ensures that your agent works within the controls you’ve set, whether it’s content filtering, protecting prompt injections, or preventing against copyright materials. So you can maintain precise control over what the agent can do and access, and everything was set up automatically when I deployed this agent. So if I come back here to our running agent, you can see that it’s returned some results. It looks like the Zava DIY is my top priority based on all of the signals that it found and looped through. So in this case, it’s worked on behalf of me using my identity and permissions to look at my calendar and surface the accounts that I should be paying attention to.
- It makes sense you’d want to have the right access controls in place because it is actually needing to look at your inbox. For example, your calendar, your data, and your file stores.
- Yeah, and this is super important to make sure that you’re building a compliant systems. Related to enterprise readiness, there are a couple of other things that I want to show that you get directly from Foundry. So in Foundry, this is my area to build and work on my agent. I have monitoring and traces. I can understand all of the conversations that might be happening, how my agent’s going about answering each questions and the overall health of my agent. Everything I need for observability is all right here. Next, there’s publishing the agent so that people can find it. So once I have my agent up and running, how do I now get this into the hands of all of my salespeople? Nobody likes building a new app, and then just hoping that everyone finds the link and bookmarks it. Well, in my case, I know that everyone in my company is using Microsoft 365 and Teams. So right here, I have a Publish button. I can take any agent deployed inside of Foundry and publish it directly to those services. This registers the agent so people can discover it and start using it right where they already work, right from Microsoft 365 on their desktop or on their phone away from the office.
- So there’s direct integration then right in Microsoft 365. In this case, in the Copilot Chat experience. And by the way, it’s also available for Microsoft Teams. Now, something also integrated with the Foundry services, Microsoft’s unified intelligence layer for AI, which helps ensure that agents are grounded in the right knowledge and also business context to keep their outputs useful and relevant. And all that goes way beyond a single source MCP server. So for example, if the agents working on your behalf, then Work IQ provides the context for how you work with the connections to your email, your calendar, your previous meetings, your Teams chat and files and more. And then you’ve got Fabric IQ, and that can be used to add context over your connected business operations. Think of things like sales data or customer records or logistics. Then you’ve got Foundry IQ, which lets you combine multiple knowledge sources for your agents, where everything from structured data sources and databases to unstructured data in your cloud stores, even images can be retrieved by agentic processes. And so Jeff, of all those different IQs that we looked at, we saw Work IQ. In that case, the agent was actually pulling from your calendar. So can we see and go deeper maybe on the rest of the intelligence layer?
- Of course, this agent has a few more tricks up its sleeve. So if we come back to the code, you’ll see that this agent actually has access to the three IQs that you just mentioned. Work IQ, Foundry IQ, and Fabric IQ. Now, based on the tools and skills I give it, let’s go back to the playground and show them in action. Again, the agent’s previous output says that I have an important meeting coming up with Zava, so I’m going to use this agent to help me get ready for this important meeting. I’m going to say, help me prepare for my upcoming meeting with Zava. Now watch what happens inside the sandbox. The agent is doing exactly the things we just described. Again, it’s checking my Work IQ to understand my correspondence, pulling in emails and Teams conversations that I’ve had with Zava. Next, it’s reaching out to Fabric IQ to pull usage data, purchasing patterns, and contract details. And it’s using Foundry IQ to search through our sales enablement materials, marketing content, to find what’s most relevant for them. Now, I’ve incorporated a few skills into this agent using the popular agent skills pattern. For example, there’s a skill defined that generates a PowerPoint presentation, another skill that creates briefing documents using Microsoft Word. So this agent came back with two file linked artifacts, a personally curated Word document for our internal team and a custom PowerPoint presentation that I can use with Zava. So I’ll go ahead and open each of these up, starting with that briefing document. You can see this has synthesized all of that contextual data retrieved from that intelligence layer, our CRM system for the relationship content and my correspondence for recent communications. It’s gone into all of the business analytics and health usage and metering, our ticketing system for support tickets. All of this is creating recommended discussion topics all into a single preparation document this agent generated. Now, if I go back, I can even show you the linked PowerPoint presentation that was generated using my other agent skill. Now, this file is actually personalized specifically for my interaction with Zava. It’s using our own company’s brand colors. You can see it’s pulled information and integrated it from Fabric IQ and Foundry IQ to give me the right talking points and relevant customer specific insights about our recent activities with Zava. It’s pulled in business operations data and included campaign metrics, including new opportunities and services that I can explore to help me build towards the next steps to take our partnership with Zava to the next level. And that’s the power of not just deploying agents, but having them run on top of the Microsoft intelligence layer, working on your behalf to access your work data, your business data, and your organizational knowledge. And it’s all integrated seamlessly with Microsoft 365.
- So now we’ve seen the agent running, we know what it can do. Now for all the developers that are watching and they’re interested in building something like this, can you explain what’s behind it and how you made it?
- Sure, so before I show you what’s behind the scenes of that more advanced agent, let’s go ahead and start with something more simple quickly here on my laptop. So for this, I’m using the Azure Developer CLI. I’ll go ahead and initialize a new project and say I want to create a sales prep agent. Now, one thing to mention, you can absolutely create chat-based agents, which are super popular. You can use any framework that you want, including things like LangGraph. And with Foundry agents, we also support emerging patterns. You’ll see we have templates to help get you going fast. So if I go ahead and choose this template, it’s going to scaffold all of the files that I need. So from here, it’s actually really straightforward. I can start debugging locally, deploy, and everything is ready to run. So this is a simple agent that I can use with a template, but there’s a lot of customization options. So we can now go ahead and go back to our advanced sales prep agent from before and look at some of what’s happening behind the scenes in the code. So you can see here, this is where I’ve defined the tools and knowledge sources. So you can see those three IQs that we walked through before. But there are some other types of skills here as well that I’m able to create and include in my coding agent patterns today. So at the core of all of this power, this agent is using the GitHub Copilot SDK. This runs a powerful agentic loop over the set of tools that I’ve defined. So when my agent was reasoning before over dozens of files, emails, and previous meetings, as well as operational and service-specific data to find relevant insights, all of this was generating informed recommendations powered by the Copilot SDK. To pull everything together, I’m using Microsoft Agent Framework. This helps me define additional pieces like middleware. So for example, here I’ve defined that if the coding agent ever tries to generate one of those documents, but it doesn’t have enough data from one of those three IQs, I want to block that because without that grounded data from all those sources, this output is almost guaranteed to be hallucinated. So these types of patterns are critical when you’re scaling deployment within an enterprise, wanting quality controls across the entire sales team, and additional guardrails and controls. Now, of course, the real power gets unlocked when I combine both these frameworks and patterns, but I host it inside of the powerful capabilities of the Foundry Agent Service.
- Okay, so in our case, we’ve published and we’ve built out two different agents. Why don’t we fast forward in time a little, one of my favorite parts of these shows, or maybe we’ve got a couple of agents running, we want to be able to monitor and manage them. What can we do there?
- Yeah, we can do all of this because it’s all running inside of Foundry. So moving back to the Foundry portal, I can manage performance costs of my entire fleet of agents in one view. So I can go ahead and look at the agent health on alerts. It looks like mine appear healthy. No alerts for me yet. I can see my estimated cost, success rates, and token usage, along with drill-in details about run volumes for our top agents. And the top and bottom agents for success rates help me see what might need attention. So you can see everything that I need to go from experimentation to production and publish across all of my end users is all right here, built-in, with full observability,
- Right, and all this is really about reducing complexity of building out and deploying your agents safely and reliably across your organization. So how can everyone who’s watching right now learn more and get started?
- Yeah, so the best way to learn is to try some of these things out for yourself. So everyone here can go to Microsoft Foundry at ai.azure.com to build your very first project. And be sure to check out github.com/microsoft-foundry. There’s a number of samples that you can try to find the SDK that you want and start coding.
- Great to have you back on, Jeff, and thank you so much for joining us today. And as always, be sure to keep it locked in here on Microsoft Mechanics, and we’ll see you again soon.
