Lori MacVittie (00:02.968) Welcome back to Pop Goes the Stack, where AI agents are coming for your workflows, starting with the ones you forgot to document. I'm Lori MacVittie, and today we're going to discuss what happens when autonomy meets entropy.I love those words. I love those words. So we've been reading about Microsoft 365's new agent ecosystem, operating system, there's a whole bunch in here. Joel's read the article, as always. Thank you, Joel. Joel Moses (00:30.702) Of course. That's what I'm here for, Lori. Lori MacVittie (00:32.738) He'll break it down for us 'cause that's a nice thing he does for us. But you know what it says is the new agent ecosystem has arrived. It's got an OS, enterprise grade controls, designed to give agents the same thing every other enterprise citizen wants: an identity, a set of permissions, a place to live, and someone to clean up after it when it goes off the rails. Joel Moses Ha ha. Lori MacVittie Yeah, so hey, they're no longer like clever features you bolt on. These are real things. They are running. you know, we know there's shadow agents, there's AI, everything's going, right? It's sitting there. So enterprises are generally like, How do we get control of this? Joel Moses Yeah. Lori MacVittie And Microsoft is offering them, Hey, here's this ecosystem of tools that you can use to actually manage them. Well, may not be the best option. So we wanted to talk to an expert, so we've got our chief digital officer, Alvina Antar. Welcome, Alvina. Alvina Antar (01:33.18) Thank you. Thanks for having me, Lori and Joel. Joel Moses (01:36.163) Excellent. Lori MacVittie (01:36.264) Absolutely. We're excited to hear your perspectives and you know what you think is going on and what we're doing and you know how this all works. So Joel, why don't you kick us off with telling us what the Joel Moses (01:48.27) Sure. Lori MacVittie article actually said. Joel Moses Yeah, well, sure. Look, it's all part of announcements that Microsoft has been making over a period of time, that are all related to AI. But they had a lot of things that they had to say out of Build 2026, one of their flagship conferences for developers, where they had a lit a litany of announcements. Some of the announcements that came out of that previously were the announcements of things like entra ID for agents, an agentic identity standard that they are proposing. Of course there are lots of other standards that are proposed for that. But in this particular Build 2026, I want to highlight one thing that occurred during the conference, and that's that when they were doing a demonstration of agent controls on main stage, amazingly, their main stage demonstration failed. And it got applause from the audience because it was actually something that should have failed, but the audience was fairly certain that it was going to succeed anyway, and so they were pleasantly surprised when it failed. So I have to give them props. It's the first time I've ever seen a main stage demo fail and fail into success. But, you know, this is Windows we're talking about, and tasks failed successfully is a thing that they like to do. Lori MacVittie (03:10.328) Yeah. Joel Moses So let me go into a couple of things that they have announced. One is Microsoft Execution Containers, which is a way to contain running code like Joel Moses (03:21.382) like AI agents into part of the Windows operating system. They have something that they've announced called Windows 365 for agents, which is a linkage for agents to use Office 365 services. They've created a Windows agent framework. They even announced something kind of interesting, they have modified the venerable Windows terminal--yeah, that's the command line--to add the ability to launch and run AI agent tasks Lori MacVittie (03:52.203) Ha ha Joel Moses and access model resources directly from the terminal. That's 0.1 if you're keeping score at home. So I don't know where that feature's gonna go, but they have definitely announced it. So I think Lori MacVittie (04:04.142) It sounded like you said Joel Moses to sum everything up, Lori MacVittie you said attacks, Joel. When you said that, attacks, Joel Moses Well, Lori MacVittie like you can launch attacks. Well yeah, of course you can. Joel Moses Of course, I mean Lori MacVittie Yeah, all right. Joel Moses this is an operating system we're talking about. Lori MacVittie Yeah, true. Joel Moses So I think it's safe to say that what Microsoft is doing here is they are trying to build a Windows centric ecosystem Joel Moses (04:20.908) that enterprises can use to get a handle on agent identities, agent runtimes, containing agent code. And if you read kind of between the lines, the future, they suggest, might lie in running models locally on each workstation that a user possesses, each device a user possesses. And you know that's certainly a viewpoint one can hold. It seems to me though, Alvina, that that might not match with the reality that a lot of enterprises face with this technology, isn't that right? Alvina Antar (04:55.174) Yeah, well thanks for giving us the summarized version Lori MacVittie/Joel Moses Ha ha ha. Alvina Antar of this article. I mean if you look at it, I mean the reality is every enterprise, I mean look at Salesforce, they're doing the same thing. They're building, you know, the operating system. Now they've got, you know, they're tied at the hip with Claude, with Claude AI and Salesforce AI Force that they just announced. And so, you know, and Microsoft is following suit and you know but for enterprises, you know, we need to remain neutral. We need to be in a position where we provide choice, right? And you know, our enterprise AI foundation that we are building and that we'll continue to build and scale, you know, needs to ensure that we've got AI experiences for every user, every persona. So, you know, we've built our own homegrown F5 AI that has access to 28 plus models. Like we need to be in a position where we can extend models. I mean every week there's a new model. Maybe Joel Moses True. Alvina Antar every day, you tell me, Joel, Lori. Lori MacVittie (06:13.87) Ha ha ha, yeah. Joel Moses (06:14.626) Depends on the week, I suppose. Lori MacVittie Mm-hmm. Yeah. Alvina Antar Right? And so we need to be in a position where we pivot, where we're able to Alvina Antar (06:23.21) also introduce our own guardrails that ensures that we're not only accessing you know models, but we're guarding and routing, you know, based on the prompt and based on the need of the model of choice. We've got, you know, enterprise foundational experiences with Gemini, with Snowflake Cortex, with you know Salesforce. And so we're running an enterprise ecosystem Joel Moses Mm-hmm. Alvina Antar with Claude, right, you know and so the need to be able to provide choice, to provide neutrality across these AI experiences, and then have the ability to have a foundation that focuses on our security our semantic layer our governance is our strategy. And then on the identity front, since you mentioned identity which is Joel Moses True. Lori MacVittie (07:21.408) Ahhh. Alvina Antar near and dear to my heart. Alvina Antar (07:23.338) You know, I mean, look at we have a new class of enterprise identity, right? I mean, it's what for over 30 years we had people and machines. And guess what? Agents break both of those buckets, right? We now have agents that they're not people, but they act like people. Joel Moses Right. Alvina Antar They do all, you know, the actions that an individual does. But the need to be in a position where that agent can now, you know, respond differently based on prompt injections. And so, which is why Microsoft, like you mentioned, introduced Entra Agent ID. We've got, Joel Moses (08:08.974) Right. Alvina Antar you know, AWS who introduced AgentCore Identity built on Cognito. We have Okta that has Agent SSO. Alvina Antar (08:19.888) You know, so the need for enterprises to truly have as part of their security fabric, agentic identities is a whole new class of enterprise identity. Joel Moses Got it. Lori MacVittie (08:32.034) Yeah, and what I hear you saying is that this package that they put together, it's nice, but there's two things going on. One, you might want to use something else, a different identity, but also that I mean no single vendor has a complete stack that solves this. Microsoft may have like all of the good pieces about containing it on the desktop, because they pretty much own desktops, but there are other pieces like you were mentioning, right, prompts and watching for injections, so guardrails and different types of protections that is not necessarily their forte, right? Joel Moses Mm-hmm. Lori MacVittie The market has continually told us for 20 years now that they want that choice of doing like best of breed. They rarely settle on like just one thing for all of their needs. It just doesn't happen. So I think that's where we're at, is they may have some good stuff here, and I think they do, especially around the desktop and managing right local, because who else is gonna do that? But there's still a lot of pieces to this platform you were talking about, this foundation that you need to get. It's not all gonna be just off the shelf, here we go. Alvina Antar (09:47.634) Yeah, I mean we're truly a hybrid multi-cloud business, right? We're running, you know, our infrastructure, our applications are running in an hybrid multi-cloud environment. And so that choice, that neutrality is key to our architecture principles. Joel Moses (10:08.13) Let's talk for a moment now, Alvina, about the rationale behind offering choice. 'Cause I think there's a number of different reasons you want to offer choice, some of them having to do with the quality or the best of breed models or pursuing the best of breed, but not always. Sometimes I imagine in an enterprise, sometimes it's about pursuing the lowest cost of ownership, right? Alvina Antar (10:32.902) Yes. It's well I mean it's mainly driven by experiences, right? And you know, as every organization scales, you know, the need to be able to ensure that we're investing in the right technology, but not creating complexity and duplicative technologies that serve our business is critical. But you know, every part of our enterprise has different, you know, applications that serve their purpose, right? We've got a whole Martech stack that serves our marketing engine. We have you know an entire sales stack that serves all of our go to market engine, right? That is we have you know our financial stack, right? I mean so we have our whole you know people stack, right, that serves, you know, everything for our people experience, right? And then we've got all the technologies that serve our product and engineering. Joel Moses Mm-hmm. Alvina Antar And it ensures that we're in a position where we can you know build our products and have the environments that allow our product and engineering employees to build and deploy at scale. And so you know all these are different technologies, right, that serve each one of our employees. And so that is exactly what drives the choice and neutrality needs is, you know, as we're serving different business functions we're essentially serving the enterprise. And while our architecture requires us to ensure that we don't have duplicative capabilities, you know that happens. I mean you think about what's happening with, you know, through the cloud era and the whole shadow IT and you know all these SaaS applications. Look at now like the agent or the agentic AI, you know, shadow AI is the new term, right? And so not only do we have core enterprise applications that run our business that are sanctioned, Alvina Antar (13:00.486) we also have shadow applications that Lori MacVittie (13:02.71) Hah. Alvina Antar are, you know, and AI is extending that beyond what we had ever seen in the SaaS cloud world. Right, and so the need to be able to have, and we use we actually use workforce AI, our own capability, to be able to govern shadow applications. Right, to be able to first and foremost, you must find Lori MacVittie (13:24.814) Find them, ha, yeah. Alvina Antar (13:25.88) and know what applications exist and then drive governance. And so that's what truly, you know, drives the need for neutrality. And then going back to the point of cost, yes, we're running a, you know, we're driving decisions around architecture based on not only providing simplicity of the experience and a seamless experience, but also an efficient experience, right? And so we don't want to continue to have duplicative capabilities that serve different parts of the business. And that is why functions like us exist, is where we can, you know, identify applications that maybe the go-to-market team has introduced that can very well serve the finance team Joel Moses Right. Alvina Antar or the people team. Lori MacVittie (14:21.358) Hmm. Alvina Antar And so, you know, providing that awareness and Alvina Antar (14:27.93) and ensuring that we're leveraging technologies across the entire enterprise, and not only leveraging and extending but fully utilizing the full force of the technology that we're investing in. Lori MacVittie (14:44.952) I think that is the first time I've ever heard anyone articulate a use case for finding shadow something other than "Well, we want to shut it down," because that's usually you want to find it Joel Moses (14:58.85) Ha ha ha. Lori MacVittie and get rid of it. And you're saying, Well, ehh, if they inter-, you know, it may be useful, we don't want redundancies. It's another way to rationalize your portfolio of AI and agents and apps. Lori MacVittie (15:11.968) And that's actually pretty cool for a discovery perspective. But I liked your points about, of course, you know, the the hybrid and choice and how many models, because I've been looking at our data for next year's State of App Strategy Report, which I have on my desktop, ha ha, data, Joel Moses (15:30.654) Ooo. Lori MacVittie and it's basically saying the same thing. Right? Enterprises are going to be Lori MacVittie (15:37.066) in this hybrid multi-cloud world and they want choice because they're actually specializing with the models, which is interesting. And we've discussed this on previous podcasts about models, right, when they generate they're trying to achieve a goal. And it turns out that depending on what model you're using, they were trained on different goal optimization. So the same prompt doesn't always work the same way Joel Moses (16:03.703) Mm-hmm. Lori MacVittie in multiple models. So you actually need that choice depending on how you're employing that AI. And that matters to agents, right? A finance agent needs accuracy and numbers. It doesn't need a story, right, like marketing does. So these are two different models it needs. Alvina Antar (16:23.088) Lori, I'm so glad you mentioned that, because if you think about it, like you know I've been in the industry almost 30 years next year. And if you look at, and in IT my entire career, and so being in a position to think about, you know, for years and years, IT, you know, the functions were looking to ensure that we contained the applications that were running wild, right, across the enterprise. Well, guess what? We cannot, I mean, we cannot be a function that says no. You know, now everyone's an engineer. I mean, think about it. What AI has created, everyone is building no code, low-code agents. We're all engineers. Joel Moses (17:09.531) Mm-hmm. Alvina Antar We're not the only enterprise engineering function that serves our internal business and ensures that we're running our business. We're in a position where we're empowering the entire enterprise to build, but build securely and build in a governed way. And so the importance of AI security is so critical Joel Moses Mm-hmm. Alvina Antar because we're not in a position where we can just find shadow applications and shut them off. Like we need to be in a position where we have the awareness and visibility and we have the proper guardrails, the proper governance, the proper identity--agentic identity--that we are embracing Joel Moses (17:58.531) Mm-hmm. Alvina Antar you know this new world that we're living in. And that we have the foundation to be able to run in this new world in a secured, seamless way. Joel Moses (18:09.484) Yeah, I think it's really practical to look at identity as kind of the linchpin of how this technology will ultimately be managed. But I do want to relate a story and then I wanna ask you a question about this to kind of close out some things. I was talking with someone from a financial services company recently and they were plagued by something that was repeatedly locking out service accounts for services that were running inside of their enterprise. And what they realized is that six months prior, one of their senior engineers had built an autonomous agent to kind of query data silos internally. And then that engineer left the company. And that agent lost its linkage to his personal enterprise credentials, but the agent was still running and because it couldn't authorize itself to the service accounts, it was locking those accounts out through repeated access. So it strikes me that identity is one thing, but also managing the lifecycle of where the agent is authenticating might also be something we have to do. What impact do you think AI is going to have on the old venerable thing of managing IT lifecycles? Alvina Antar (19:23.292) Joel, it's I mean, look at onboarding, right? I mean, when a new hire shows up, we give them an identity, a laptop, access, you know, based on their role, role-based access. The agent is equivalent to that, right? We register it, we bind it to an owner, we define what its purpose is and where it should run. But the whole model breaks, right, when it relates to permissions. Like the lazy path is to give the agent the permissions that the person that built it Joel Moses (19:59.681) Right. Alvina Antar who launched it, you know, has and inherits. But if you look at you know agentic permissions, the new model you need to start at zero like to time box it and Lori MacVittie Mm-hmm. Alvina Antar ensure that you have human approval for their actions and that Alvina Antar (20:20.996) the insight for a human you know for detection and response you know it was managed you know through our prior lifecycle management Joel Moses Right. Alvina Antar processes. But for an agent running at machine speed, that whole gap is the whole breach. Right? And so the need to ensure that we really think about how we're, you know, how do agents resign? You know, what is Lori MacVittie Ha ha ha. Alvina Antar What is the offboarding of an agent? Lori MacVittie (20:54.423) No. Alvina Antar Do they don't hand in a badge? They don't like Joel Moses (20:59.454) Yeah. Right. Alvina Antar they just keep running with their standing access. And if you look at, you know, the majority of breaches are orphan service accounts. Joel Moses (21:10.508) Yeah. Yeah. Lori MacVittie (21:10.562) Yeah. Mm-hmm. Alvina Antar Right? Or just service accounts that are unmanaged. And Joel Moses (21:14.732) I think, I agree, I think structured onboarding and structured offboarding for agents is something that probably needs to happen. Just like you onboard an employee, you probably should onboard an agent in a very consistent fashion that matches the lifecycle management of the rest of the accounts that you manage. That's interesting stuff. Lori MacVittie Yeah. Joel Moses Yeah, Lori MacVittie (21:36.718) But we gotta wind it up because we're Joel Moses I know. Lori MacVittie we're running up against time and I know we could talk for hours, Joel, but you know. Come on, any last words of advice for enterprise listeners? Like what should they take away? You know, should they just Joel Moses (21:52.877) Yeah. Lori MacVittie go out and buy this? Should they look it over, build their own what's going on? Joel Moses (21:57.516) Yeah, well I've got one takeaway, and that's that ecosystems are great things. They give you a starting point, but they invariably run short when you have to customize, seek alternatives, or seek cost containment. And so I'm very heartened that Microsoft is doing all of these tools. I wish them luck. I hope they keep their eye on interoperability. But enterprises, the ground truth is that we're gonna be each building our own way through this. Lori MacVittie (22:28.648) Awesome. Any last advice, Alvina, for your peers out there? Alvina Antar (22:33.062) Well, I couldn't agree more. And you know, choice and neutrality in this hybrid multi-cloud environment is key. And the need to be able to focus on building the enterprise AI foundation that provides that choice. And be in a position where you've got a shared foundation focused on security and governance that allow for the enterprise to have the neutrality and continue to innovate at the pace that AI requires is key to success in this new world. Lori MacVittie (23:17.838) Absolutely. Absolutely. Great advice. Great talk. Loved it. But that's all the time we have for this week. So, you know, hey, that's a wrap. Don't forget to subscribe, please, before your automation starts automating new problems for you to automate. And new problems...this is a recursive problem. So hey, subscribe. Prevent this.