Satya Nadella Explains The Future of AI-Powered Coding and Multi-Agent Systems

The IT backlog or the tech debt that exists in the world broadly, you need more software built such that it can address the world's needs.

Copilot tuning is it's not just about using intelligence, but it's about tuning intelligence to amplify a given firm's ability to produce the products and services that they anyway are doing and doing it with much more leverage. Your point is absolutely right. We do need these social systems and a language. That's why I think the fact that it's all natural language, the work artifacts, that I think becomes the interface.

Yeah, and know the way I, Tiffany, come at this is if I look at the demand side of software development, the reality is the IT backlog or the tech debt that exists in the world broadly—not just in the narrowly defined tech industry, right? When you look at a small business in Indonesia to a large multinational trying to in manufacturing in Poland, and you sort of look at the world—then you need more software built such that it can address the world's needs. So if you say the demand is there, and there's a massive deficit of actually supply, so therefore, all of these dev tools, whether it was just simple code completions from a few years ago to Copilot Chat to Copilot agent mode or even what we launched today with Copilot coding agent, all are really about helping software developers and software development teams stay more in the flow, working down that tech debt more, producing the tech solutions that the world needs. And then the other piece, which is this Copilot tuning, to me is again more like another tool for any software engineer in any company to be able to then turn the particular firm's competitive advantage—which is what is it, the knowledge they have as a firm, the documents they have or the databases they have—but then to use the latest model plus that data to turn it into sort of the competitive advantage for that firm. That's what this Copilot tuning is. It's not just about using intelligence, but it's about tuning intelligence to amplify a given firm's ability to produce the products and services that they anyway are doing and doing it with much more leverage.

That's correct. Absolutely. Well said.

Yeah, I mean that's the real question in front of us is what's the social system of trust that exists? And I think the social system of trust would be very similar in some sense to how even we as co-workers, when you're working on a project, we're able to look at the work artifacts, we're able to inspect what we are doing in our work to move things forward. So even like the coding agent, for example, I love when even I assign a PR or an issue to it, it opens a PR in its own branch. The thing that I go to is the issues log, right? So it's actually committing all of the session log where it's committing all these draft commits, and I'm looking at it and say, "Wow, every step of the way, it's like uncannily watching someone else code away and know what they're doing," right? So in some sense, your point is absolutely right. We do need these social systems and a language. That's why I think the fact that it's all natural language, the work artifacts, that I think becomes the interface. In fact, even the markdown file for agents is the interface between us and an agent. And so I do feel, whether it's in Teams or whether it's in GitHub or what have you, we will have to basically build out these things which are the interchanges where humans and agents can essentially, quite frankly, verify each other's work.