Building more than just an agent harness

**Ryan Donovan** (0:07)
Hello, and welcome to The Stack Overflow Podcast, a place to talk all things software and technology.
I'm your host, Ryan Donovan, and I'm recording today from Microsoft Build. My guest today is Jay Parikh, the VP of AI Core here. We're going to be talking about everything that it takes to run and build AI agents and tiny topic, right? Little itty bitty. So, welcome to the show.

**Jay Parikh** (0:31)
Thank you.

**Ryan Donovan** (0:32)
So before we get into the topic, let's get a little bit about yourself. How did you get into software and technology?

**Jay Parikh** (0:38)
Yeah. Well, it's been a really long time. I've just always been a builder. We all have that journey, have spent all of my career building infrastructure, software products, scaling systems, and startups.
So it's just been a wild time. I've seen the scale of the Internet, was at Facebook for a long time, scaling engineering there and serving billions of people, cloud security, now AI and platforms and tools for builders who are building for this agentic era now.

**Ryan Donovan** (1:13)
Right. Definitely things have shifted, right?

**Jay Parikh** (1:15)
Yes.

**Ryan Donovan** (1:16)
Faster and faster. So we're talking AI agents at scale, and Microsoft obviously deals with a lot of scale. Before we talk about some of the things announced today, what does it take to run and build AI agents for the enterprise at scale?

**Jay Parikh** (1:31)
Yeah. So one of the, I guess, many things that we have announced, one of the core things that we've rolled out, that we put together today, is how you put together the different components and how you think about what it's going to take to be successful with agents at scale in the enterprise. So there's obviously a lot of buzz, energy kind of innovation that's happening around agents.
But a lot of where I spend my time talking to customers, working with customers, influencing, and working with our teams of how we build our products is the myriad of use cases that enterprise customers have. And it's trying to assemble these things piece by piece, really hard. They're all changing out underneath you all the time.

**Ryan Donovan** (2:17)
But it's building their own stuff.

**Jay Parikh** (2:18)
And so showing them this, in some ways, paved path or this like, hey, this is how these components connect to each other, and this is how you can run this loop. So very simply, you start off building your agents, and you can do that in, say, VS Code or the new GitHub app, as we've demonstrated today. So you can get together your assets that you need for an agent, different frameworks, you write your evals, you're tuning that logic that you want for your business process, right? Whatever that may be. Could be a compliance thing, it could be an onboarding thing for a new employee, it could be a sales agent, et cetera.
Then you need to connect that because even though you may pick a model, you do need to connect that to your enterprise data to make that agent succeed more often and to deliver that ROI, that correctness, that quality that is required with your enterprise environment. So connecting that into Microsoft IQ where we help bring that data and surface it and connecting it to your agent, fabric, these types of tools, Web IQ, being able to ground in the real world data that's happening in the Internet, then you need to be able to run these agents. And there's three aspects to running this. One is building and running the agents in a purpose-built platform called Foundry. And we'll come back to that, but this is a new platform that has been out for a little over a year now that gives you model choice, but it gives you a bunch of other things that you need from a primitive perspective to run those agents, to do harder things and to run them for longer periods of time. Then, obviously, in the enterprise, but it doesn't matter if you're a big enterprise or a small enterprise, you have to secure these agents, right? You have to make sure that they're not going off the rails, that you can observe them, that you can manage them, right? Because you want to understand how they're accessing data. What are they allowed to? What are they not allowed to access? So securing that with the Microsoft Security Stack. And then finally, putting those agents in the flow of work, because many agents still need review, interaction, approval from humans. So putting that agent in the flow of work in M365, or in Teams, or in email, like we need to be able to do that. And that's what I think is really exciting when we talk to our customers, our ability to put this build, deploy, run, and all the different components and products together into this end-to-end system that has that enterprise grade, trust, security, and it also is in a loop where as the system, as these agents run longer, the system can continually improve. Yeah.