Microsoft's Jay Parikh: Agents will get rid of your coding backlog

SAN FRANCISCO — Jay Parikh played a key role building the infrastructure that allowed Facebook to grow from a fun site to share party photos and catch up with old friends into a global megacorporation that reached 3.5 billion daily users in the last quarter. Now at Microsoft, his mission is to help a different kind of giant rethink the way it builds software.

Parikh, executive vice president of Microsoft's CoreAI group, leads a relatively new organization that is overhauling Microsoft's software development strategy — both for internal tools and for external customers — around the new capabilities that large-language models have brought to the command line. "This is leading to a new AI-first app stack — one with new UI/UX patterns, runtimes to build with agents, orchestrate multiple agents, and a reimagined management and observability layer," CEO Satya Nadella said in a memo announcing the creation of the group in January.

"Having been in enterprise for a long time and been a developer, just to think about how things were done maybe even, like, five or six years ago, versus what I see [from] the most advanced teams within Microsoft and how they're really putting this capability in today, they're pushing it to the limit," Parikh said in an interview with Runtime this week at Microsoft's Ignite conference. While the broader merits of the generative AI revolution are still up for debate, there's no question that the software-development process will never be the same.

Parkih discussed the strategy behind the CoreAI group, how agents are already making possible things that mere copilots could never imagine, and how Microsoft thinks about the overlap between the tools it builds and uses internally to the products it builds for customers.

This interview has been edited and condensed for clarity.

Clearly you were brought in to bring about some sort of change to the engineering organization, and I'm wondering if you could tell me a little bit about what your mandate was, and then give me an update on how it's gone.

I joined Microsoft at the end of October, and then I spent a couple months ramping up and meeting people and learning as much as I could, as fast as I could. In January of this year we formed the CoreAI team, and then in May of this year at Microsoft Build the conference, we put together and explained our overall strategy for the focus of CoreAI, which I can expand upon in a minute.

And then at GitHub Universe [last month], we came back to basically the top part of that product vision that we outlined at Build around reinventing and rethinking everything for AI-powered tools across the entire software development life cycle. Today, here at the Ignite conference, we're entirely focused on how that relates to the enterprise, and specifically the application platform that these modern AI applications — or, more and more these days, agents — are built, deployed and operated on.

We mostly have these AI agents running in the cloud, but we are building technology so that you can bring them to the edge for certain use cases. Now, I'd say we're still early in that journey from a customer perspective, but we want to have that capability so that enterprises who are deploying this stuff at rapid scale have that option.

Can you point to some specific things that have changed in the software development life cycle with respect to some of these AI tools?

It's super fascinating to see this evolution or revolution, whatever you want to call it. Having been in enterprise for a long time and been a developer, just to think about how things were done maybe even, like, five or six years ago, versus what I see [from] the most advanced teams within Microsoft and how they're really putting this capability in today, they're pushing it to the limit.

There's a set of people out there … they're not using it, they're still kind of doing things the old way, and I think they are at risk of being left behind. They might be just doing a little bit of code completion stuff, but they're not really harnessing the kind of creative power and the throughput that you get with coding agents.

And when I say coding agents, it's not just the thing that generates the code. The way I see the more advanced engineers, or the more experienced engineers on AI use this technology, is they're spending a lot of time thinking about context, and setting up that architecture of what they want the outcome, the output, to look like.

They're masters of this, right? There's a craft to it. There's a structure to it. So there's an objective part of it, there's a subjective part of this, and then they're able to take that, harness that, and kind of shove it into a team of agents; not just giving it a task, like we demo.

A lot of our demos are like, "hey, here's a GitHub issue." We assign it to Copilot, it does that task, [and] it's amazing because maybe it's a bunch of work I don't even want to do as a developer. And now I can assign and get this buddy, Copilot, to do this stuff. I will verify it, I'll look at the code, I'll make sure it's right and it's working and been tested, and it's secure before I push it to production. But now I can offload these things, and I can save my time for the harder, more creative things to focus on as a developer.

What I found more interesting, believe it or not, was the fact that within, like, two months later, that team then basically took their entire backlog of features, bugs, everything they had in the backlog, and that backlog was drained to zero things. That is just incredible; like, who talks about having zero backlog?

But the more advanced teams and engineers are actually able to take that creation and assign it to a team of agents, and have these agents work together to accomplish a much more complex task. They will then use agents to help them verify too; they're in the loop, they're verifying things, they're testing things, they're double-checking things, and then they're getting the assistance from other agents to deploy things right and to fix things in production.

So they think agent-first in everything they do and not, "okay, let me see how many things I can type myself or click ops myself. Let me see what I can get an agent to do for me." And that's where I see the teams [that] are really pushing the limits of what these tools and systems can do today.

Singapore Airlines had this project where they wanted to revamp some part of their consumer application. That project originally was scoped for 11 weeks. They were able to use GitHub Copilot and deliver that project in five weeks, so a very big reduction in time to market for them.

What I found more interesting, believe it or not, was the fact that within, like, two months later, that team then basically took their entire backlog of features, bugs, everything they had in the backlog, and that backlog was drained to zero things. That is just incredible; like, who talks about having zero backlog? That is an outcome that I don't think we've ever been able to realize, and now, with AI and being able to have the right training, the right adoption, the right incentives for our teams, that could be something that really does translate into reducing technical debt, having more secure applications, delivering more value, more features, more capabilities, to the customers.

How strict is Microsoft about "golden paths" and things like that for internal developers? Are developers given a real strict template like, "this is the way we do things," or do you allow some leeway?

First and foremost, we have a set of security guidelines that we adhere to. That's non-negotiable. From there, there are lots of things that we experiment with in terms of tools we build internally on top of these AIs that help us, and they may land in another product, like GitHub Copilot, or something else. We do obviously try out lots of other tools from other companies and open source and whatnot, and then ultimately, anything that ends up in a customer's hands or in production goes through a very tight set of tests and checks.

We use lots of different tools, we make lots of different tools. We have this thing, for example, called SRE Agent, and that was a tool that we developed internally to help our engineers manage all of this massive scale operational work. Now we are building that for our own internal productivity and governance and just overall effectiveness, but then we'll turn around and offer that to customers right at some point.

That was done by an engineering team to solve a set of problems for a set of neighboring teams, and more and more Azure teams are using it, and then it's helping to make that product better. And then we'll add a bunch of controls and visibility and other things, then turn that into a product to offer to customers.

Editor's note: Microsoft provided Runtime with travel and accommodations to attend Ignite 2025.